Incorporating algorithmic uncertainty into a clinical machine deep learning algorithm for urgent head CTs

Abstract

Machine learning (ML) algorithms to detect critical findings on head CTs may expedite patient management. Most ML algorithms for diagnostic imaging analysis utilize dichotomous classifications to determine whether a specific abnormality is present. However, imaging findings may be indeterminate, and algorithmic inferences may have substantial uncertainty. We incorporated awareness of uncertainty into an ML algorithm that detects intracranial hemorrhage or other urgent intracranial abnormalities and evaluated prospectively identified, 1000 consecutive noncontrast head CTs assigned to Emergency Department Neuroradiology for interpretation. The algorithm classified the scans into high (IC+) and low (IC-) probabilities for intracranial hemorrhage or other urgent abnormalities. All other cases were designated as No Prediction (NP) by the algorithm. The positive predictive value for IC+ cases (N = 103) was 0.91 (CI: 0.84–0.96), and the negative predictive value for IC- cases (N = 729) was 0.94 (0.91–0.96). Admission, neurosurgical intervention, and 30-day mortality rates for IC+ was 75% (63–84), 35% (24–47), and 10% (4–20), compared to 43% (40–47), 4% (3–6), and 3% (2–5) for IC-. There were 168 NP cases, of which 32% had intracranial hemorrhage or other urgent abnormalities, 31% had artifacts and postoperative changes, and 29% had no abnormalities. An ML algorithm incorporating uncertainty classified most head CTs into clinically relevant groups with high predictive values and may help accelerate the management of patients with intracranial hemorrhage or other urgent intracranial abnormalities.

Background

A patient with an urgent neurological abnormality requires prompt imaging. If the etiology is hemorrhage, immediate evacuation or treatment of an aneurysm may be lifesaving. If the cause is an ischemic stroke from a large vessel occlusion, thrombectomy can result in a favorable outcome. There are many additional causes of urgent neurological deficits including masses, infections, inflammatory diseases, and acute hydrocephalus for which prompt intervention could markedly decrease morbidity and mortality. Rapid identification of such a finding is critical for triaging patients to appropriate management. CT is commonly the first imaging method employed in this situation, and computer-based image evaluation tools that rapidly classify and triage CT scans would be valuable.

The application of artificial intelligence (AI) in neuroimaging has expanded with the potential to accelerate the accurate diagnosis of intracranial hemorrhage or other urgent intracranial abnormalities. Many machine-learning (ML) algorithms rely on dichotomous classification schemes where an algorithm determines the presence or absence of specific intracranial abnormalities such as intracranial hemorrhage, large vessel occlusion, or metastasis [1–4]. These algorithms have demonstrated high accuracy in identifying intracranial abnormalities. For instance, more recently reported algorithms were shown to detect intracranial hemorrhages with specificities of 0.93–0.95, sensitivities of 0.87–0.96, positive predictive values of 0.71–0.96, and negative predictive values of 0.95–0.95 [4–6]. However, abnormal findings on neuroimaging are not always definitive. For instance, a streak artifact on CT may mimic intracranial hemorrhage. Uncertainties can also be introduced at different stages in the development of ML algorithms for various reasons including suboptimal input data, weak supervision, as well as inter- and intra-observer variability [7]. With the expansion of real-life applications of various ML algorithms and the growing need for safety, the recognition and appropriate handling of uncertainty is becoming critical [8, 9].

Clinical ambiguity is also common. For instance, stroke symptoms including a facial droop and speech difficulty may not only result from early ischemia but also from acute intracranial hemorrhage or a mass lesion. An appropriate neurological assessment may not be feasible in patients who are obtunded from various causes such as ischemia, infection, and inflammation. The clinical issue at hand is therefore complex, and more than one ML algorithm may be needed for adequate characterization of the head CT, including an algorithm that rapidly identifies a major abnormality such as hemorrhage, mass lesions, and other stroke mimics as well as another algorithm that quickly assesses the likelihood of ischemic injury.

In this study, our aim was to evaluate an ML tool that classifies non-contrast CT scans into three categories based on the probability of the presence of an intracranial hemorrhage or other urgent intracranial abnormality other than acute ischemia. We assessed the impact of this classification in predicting clinical outcomes including hospital admission, neurosurgical intervention, and 30-day mortality.

Materials and methods

Deep learning algorithm

The algorithm employed was derived from an intracranial hemorrhage detection and subtype classification system first created by finetuning four ImageNet-pretrained—deep convolutional neural networks (DCNNs)—VGG16 [10], ResNet-50 [11], Inception-v3 [12], and Inception-ResNet-v2 [13], with ICH training data obtained from the imaging archives of our institution [14]. Specifically, 5 mm thick two-dimensional (2D) axial images from 904 non-contrast head CTs were labeled by 5 subspecialty-trained US board-certified neuroradiologists as to the absence or presence of intracranial hemorrhage as well as all identifiable subtypes (intraparenchymal, intraventricular, subdural, epidural, subarachnoid). These studies were randomly divided into training and validation sets and additional retrospectively and prospectively collected data sets were used for testing. Finetuning of the DCNNs with this labeled data occurred after the last fully-connected layers were replaced with three consecutive layers containing a global average pooling (GAP) [15] layer, a fully-connected layer, and an element-wise sigmoid layer. All models were optimized using a mini-batch stochastic gradient descent with Nesterov momentum [16] with a batch size of 64 to maximize GPU utilization. We used a weight decay of 5x10^-5 and a base learning rate of 0.001, decayed by 0.1 three times when the validation loss plateaus. Computations were conducted on an NVIDIA DevBox equipped with four TITAN X GPUs with 12GB of memory per GPU, and all deep learning models were implemented using Keras (version 2.1.2, http://keras.io) with a Tensorflow [17] backend (version 1.3.0). The model provided a probability on a slice-by-slice basis of the presence of intraparenchymal hemorrhage, interventricular hemorrhage, subdural hematoma, epidural hematoma, and subarachnoid hemorrhage. This multi-label classification task was reformulated into a binary classification of intracranial hemorrhage as positive if one or more of the subtype outputs are positive and negative if not. The binary cross entropy loss function was weighted by the ratios of positive and negative instances for each class label, in a similar fashion as described previously [18, 19]. An ensemble of the four models was created using unweighted averaging such that the final probability is defined as an average of probabilities predicted by the four models. We also applied the Z directional moving average window method to determine the results of the target image by considering the results of the images above and below the Z-axis. This method is more accurate than the 2D method and has better memory usage and algorithm speed than the 3D method. The model assigned for each slice of a CT scan, a probability of 0 to 1 for the presence of intracranial hemorrhage. Complete details are described by Lee et al. [14].

Adapting the algorithm for clinical use

The algorithm underwent a preliminary evaluation using ED neuroradiology head CTs acquired over a period of one month. Review of the results revealed that the cases that the algorithm categorized as IC+ included not only those with ICH but also other significant intracranial abnormalities such as diffuse cerebral edema and mass effect. We deduced that this was a form of transfer learning and a valuable feature of the algorithm. We also observed that when a study had 3 or more slices with probabilities of at least 0.9, ICH or other significant intracranial abnormality was very likely. We also noted that if a study contained just 1 or no images with probability greater or equal to 0.6, ICH or other abnormality was very unlikely. For our prospective study we classified studies with 3 or more slices with probabilities of 0.9 or greater as IC+; studies with all slices except one having probabilities of 0.6 or less were classified as IC-. All other studies were given the designation of NP indicating no prediction by the algorithm.

Imaging test data collection

This study was approved by the institutional review board (Mass General Brigham IRB Protocol #: 2019P002147, Expiration Date: 09/12/25). Prospectively, we identified 1000 consecutive non-contrast head CTs assigned for review by the emergency department neuroradiology service. Most of the studies were non-contrast head CT, but some were performed in conjunction with CT angiography. Regardless of the type of the study, only the axial, 5-mm, noncontrast, standard reconstruction kernel, head CT images were evaluated. Most of the studies (N = 953) were acquired in the emergency department or inpatient CT scanners (GE Discovery CT750 HD, GE LightSpeed VCT, Siemens SOMATOM Definition Edge). The remaining 47 studies were performed on the outpatient scanners (Philips IQon -Spectral CT, GE Discovery CT750 HD).

The clinical information was derived from electronic medical records. In addition to patients’ demographic data, the rates of admission, neurosurgical intervention, and 30-day mortality were also collected. Neurosurgical interventions included craniotomy, craniectomy, cranioplasty, ventricular catheter placement or removal, biopsy, and interventional neuroradiology procedures. If a patient had multiple imaging exams during the same admission, it was counted as a single admission. There were 4 patients who were discharged and re-admitted subsequently during the data collection. Their admissions were counted as two separate admissions.

Imaging data analysis and clinical information

Each CT scan classified by the algorithm was compared to the formal clinical interpretation by a neuroradiologist. All the neuroradiologist in the study were board-certified with Certificates of Added Qualification (CAQ). The clinical interpretations were made without the knowledge of the ML algorithm results. The presence of intracranial hemorrhage or other urgent intracranial abnormalities was recorded and compared with the algorithm results.

Statistical analysis

Descriptive summaries were computed for the overall cohort. Continuous variables were summarized as the median and interquartile range (IQR: 25^th and 75^th quantile) and categorical variables as percentages.

Separate generalized linear regression models were constructed to quantify the association between classification (IC+, IC-, NP) and the presence of each imaging finding (acute intracranial hemorrhage, mass lesion, artifact, post-op changes, acute/subacute infarcts, encephalomalacia, miscellaneous, and normal imaging) and each outcome (hospital admission, neurosurgical intervention, and 30-day mortality). The parameters of these models were estimated using generalized estimating equations, with an independence correlation structure (GEE-I), to account for the possibility of multiple scans per subject. Linear combinations of parameter estimates were computed to summarize the prevalence of each imaging finding and the occurrence of each outcome by classification group as well as all pairwise comparisons (outcomes only) between classification groups along with their associated confidence intervals and p-values. Similarly, GEE-I estimates were constructed to estimate predictive values while acknowledging the possibility of multiple admissions and scans per subject. All analyses were performed using R 4.1.1 (R: the R project for statistical computing; www.r-project.org) and the geepack library [20].

Results

One thousand consecutive head CT scans were performed on 857 patients. Some patients had multiple studies during the data collection period. Of these 857 subjects, 761 had a single scan, while 66 had two scans, and 33 had three or more scans. There were 423 female (49.4%) and 434 male (50.6%) patients. The median age was 65 years (IQR: 50–77). The overall hospital admission rate was 47.1% (CI: 43.8–50.4) which is higher than the ~30% admission rate of all patients evaluated in our institution’s emergency department. Neurosurgical interventions, including craniotomies and endovascular procedures, were performed in 7.7% (6.1–9.7) and the 30-day mortality was 4.4% (3.2–6.0).

Of the 1000 scans, 10.3% were classified by the algorithm as IC+ (high probability of an intracranial hemorrhage or other urgent intracranial abnormality), and 72.9% were classified as IC- (low probability of an ICH or other urgent intracranial abnormality). Studies classified as No Prediction (NP) constituted 16.8% of the 1000 scans. Table 1 lists the imaging findings tabulated for IC+ and IC- cases.

[Figure omitted. See PDF.]

The number (percentage, 95% confidence interval) of major intracranial abnormalities for studies classified as IC+ and IC-. Intracranial hemorrhage includes acute hemorrhage associated with aneurysmal rupture, posttraumatic, hemorrhagic infarcts, hemorrhagic mass, and/or postoperative changes. Urgent findings associated with mass/mass effect includes herniation, ventricular effacement, and/or ventricular entrapment from various etiologies including neoplasm, infection, inflammation, and cerebral edema. Postoperative changes include findings such as resection cavity or pneumocephalus from craniotomy without acute intracranial hemorrhage or substantial mass effect. Acute/subacute infarcts include non-hemorrhagic ischemic infarction. Miscellaneous includes ventriculomegaly, chronic subdural collections, and hygromas. All estimates, and p-values, are based on separate generalized linear models whose parameters were estimated using generalized estimating equations using an independent correlation structure.

The algorithm classified 103 cases as IC+. Examples of IC+ scans are shown in Fig 1. Over 80% of these patients had intracranial hemorrhages, either primary or secondary to underlying lesions identified on the non-contrast CT or associated examinations.

[Figure omitted. See PDF.]

A: Examples of IC+ cases with acute intracranial hemorrhages. The left panel shows a large intraparenchymal hematoma centered in the left basal ganglia (arrow) as well as intraventricular hemorrhages within the bilateral occipital horns (arrowhead). The middle panel shows a large left holohemispheric subdural hematoma with rightward subfalcine herniation. The right panel shows multifocal, hemorrhagic lung cancer metastases. B: Examples of nonhemorrhagic IC+ cases. The left panel shows a large area of confluent hypoattenuation involving the right hemisphere, which was subsequently found to be toxoplasmosis in a patient with human immunodeficiency virus (HIV) and acquired immunodeficiency syndrome (AIDS). The middle panel shows diffuse reversal of gray-white differentiation as well as partial effacement of the ventricles, most consistent with diffuse cerebral edema from a severe hypoxic-ischemic injury. For instance, the left caudate head has an abnormally hypodense appearance compared to the adjacent internal capsule. The right panel shows a large, left hemispheric, Spetzler Martin grade 5 arteriovenous malformation. C: Examples of false positive IC+ cases. The left panel shows streak artifact from metallic hardware. The middle panel shows encephalomalacia of the right frontal lobe. The right panel shows an old right posterior cerebral artery territory infarction.

Non-hemorrhagic masses or lesions causing substantial mass effect were the second most common findings in over 10% of cases. Several cases with urgent pathology including diffuse cerebral edema and large vascular malformation were also classified as IC+. There were 9 IC+ false positives. There was only a single normal study that was classified as IC+ with the erroneous classification thought to be secondary to residual intracranial vascular enhancement from an earlier exam performed with intravenous contrast.

Among the IC+ cases, there were 94 true positive and 9 false positive cases for major intracranial abnormality, resulting in a positive predictive value 0.91 (0.84–0.96). Patients with IC+ scans had a very high rate of admission, 74.6% (62.5–83.8). The rates of neurosurgical intervention, 34.9% (24.2–47.4), and 30-day mortality, 9.5% (4.3–19.6), were also high. These outcomes were significantly greater in IC+ patients compared to IC- patients (Table 2).

[Figure omitted. See PDF.]

The findings on the 9 false positive IC+ scans are listed in Table 3. The false positive cases included 4 with artifacts, 4 with large areas of encephalomalacia and one from a patient that received intravenous contrast during an earlier exam.

[Figure omitted. See PDF.]

Most patients (N = 729) had scans classified as IC-, and examples are shown in Fig 2. The negative predictive value was 0.94 (0.91–0.96).

[Figure omitted. See PDF.]

A: Examples of true negative IC- cases. The left panel shows a normal noncontrast head CT. The right panel shows small old infarcts in the left basal ganglia. B: Examples of false negative IC- cases. The left panel shows a small isodense subdural hematoma along the left frontal convexity. The right panel shows trace subarachnoid hemorrhage along the medial aspect of the left frontal lobe.

The rates of admission, neurosurgical intervention, and 30-day mortality for the IC- group were substantially lower than those in the IC+ group (Table 1). There were 686 true negative cases, and they had no abnormal findings or small chronic abnormalities (Fig 2). There were 43 scans that were classified as IC- that were false negatives (Table 4). Nearly all, 39 of 41, of the false negative scans had no mass effect associated with intraparenchymal or extra-axial hemorrhages (Fig 2). The remaining 2 false-negative cases had mass lesions in the cerebellum with partial effacement of the fourth ventricles.

[Figure omitted. See PDF.]

The NP patient group included 168 scans (Fig 3). The largest portion of the scans was normal (n = 49, 29.2%, 22.7–36.7) followed by those with acute intracranial hemorrhage (n = 41, 24.4%, 17.7–32.7). Compared to the IC+ and IC- groups, the NP group demonstrated the highest proportion of scans with artifacts (n = 19, 11.3%, 7.3–17.1) as well as postoperative changes that include the presence of a resection cavity or pneumocephalus from craniotomy without acute intracranial hemorrhage or substantial mass effect (n = 33, 19.6%, 14.1–26.6). Other scans include those with non-hemorrhagic mass/mass effect with urgent findings (n = 7, 4.2%, 2.0–8.4), non-hemorrhagic mass/mass effect without urgent findings (n = 6, 3.6%, 1.6–7.8), and miscellaneous findings such as ventriculomegly, chronic subdural collections, and hygromas (n = 2, 1.2%, 0.3–4.6).

[Figure omitted. See PDF.]

A: Trace acute subarachnoid hemorrhage along the right anterior cingulate gyrus. B: Noncontrast head CT degraded by motion artifact. C: Normal noncontrast head CT.

The rates of admission, neurosurgical intervention, and 30-day mortality were 53.4% (44.9–61.7), 12.0% (7.5–18.7), and 9.0% (5.2–15.2), respectively. The rates of admission and neurosurgical intervention were significantly higher for the IC+ cohort compared to NP with the odds ratios of 2.57 (1.32–4.97, P = 0.005) and 3.92 (1.88–8.19, P <0.001), respectively. When compared to the IC- cohort, the NP cohort had significantly higher rates of admission and neurosurgical intervention with the odds ratios of 1.50 (1.03–2.18, P = 0.033) and 3.03 (1.59–5.77, P < 0.001), respectively. The rate of 30-day mortality of the NP was not significantly different compared to the rate for the IC+ cohort (odds ratio 1.06 (0.38–2.97), P = 0.910) but was significantly higher than the rate for the IC- cohort (odds ratio 3.22 (1.54–6.76), P = 0.002).

There were 33 patients who had an acute, subacute, or age-indeterminate ischemic infarction. Twenty-eight patients were classified as IC- and 5 patients were classified as NP group.

Discussion

Diagnostic ambiguity is common when a patient is evaluated for a new neurological symptom or after an event such as trauma. For example, symptoms such as a facial droop and speech difficulty may result from ischemia, hemorrhage, mass lesion and other pathologies. Even greater ambiguity occurs in patients who are unable to respond. Imaging is effective in narrowing the differential diagnosis, but imaging and image interpretation are also imperfect. Since time is of the essence, ML algorithms that quickly identify imaging abnormalities are promising new tools, but their inferences also have uncertainties that must be recognized and managed.

Following the suggestions by Kompa and colleagues [9], we incorporated uncertainty into an ML algorithm that analyzes head CT scans. The algorithmic system correctly identifies IC+ and excludes IC- scans with intracranial hemorrhage or other urgent intracranial abnormalities in most patients. The studies with high algorithmic uncertainty were placed in the NP group. This is a system that has utility as a prioritization tool in busy settings and may be especially useful to clinicians when radiologists are not available.

Transfer learning is routinely used in machine learning. In transfer learning, a network trained to solve one problem using one dataset serves as the starting point for training the network for a related problem using a completely different dataset. The first dataset provides domain adaptation and cuts down the training time for the second problem. By exploiting what has been learned in one setting helps to improve generalization in another setting. Our use of a network trained for ICH to detect other abnormalities represents a form of transfer learning that works by learning features that are common to both ICH and other abnormalities. Most likely our trained network learned certain features from the base dataset and inductively transferred them to other brain abnormalities. Our network performs well for ICH detection, but the process of training narrowed the model bias in a beneficial way where it became useful for other intracranial abnormalities such as tumors and other hemorrhage-like conditions. While our network’s performance on non-ICH conditions can be further improved by subjecting it to more specific training, its performance even without such training is remarkable.

IC+ scans

The IC+ cohort comprised about 10% of the test group and had a >90% positive predictive value for an intracranial hemorrhage or other urgent intracranial abnormality. Moreover, three-fourths of this group were admitted, over a third had a neurosurgical intervention and nearly a tenth died within 30 days. It is evident that an IC+ classification is a powerful biomarker, and those patients should be promptly assessed clinically, and their scans immediately reviewed by an imaging expert. In a setting where a radiologist is not available, an IC+ classification provides an important alert to caregivers and can prompt additional action to manage the patient. It is noteworthy that even the false positive IC+ scans had imaging abnormalities, mostly encephalomalacia or artifacts.

IC- scans

The IC- cohort had a 94% negative predictive value for intracranial hemorrhage or other urgent intracranial abnormality. This classification identified patients with significantly lower rates of admission, neurosurgical intervention, and 30-day mortality. Nearly all the false-negative cases within this cohort were due to a small amount of intracranial blood. While IC+ cases have a higher probability of more urgent findings, the presence of hemorrhage in a subset of IC- cases suggests that the IC- cases should also be reviewed by trained interpreters in a timely manner to identify small bleeds.

NP scans

The patients whose scans were placed into the uncertain NP cohort had interesting characteristics. Of the 168 NP scans, 31% had image artifacts or post-operative changes and 29% were normal. However, nearly a third had intracranial hemorrhage or other urgent intracranial abnormalities. Thus, immediate review of these scans by a radiologist or trained interpreter of head CTs is warranted for these scans, making the reviewer aware of the 1 in 3 chances of the presence of an intracranial hemorrhage or other urgent intracranial pathology.

Early ischemic stroke

The ML algorithm does not identify ischemic injury. Not a single case of ischemic infarction was classified as IC+. Acute, subacute, and age-indeterminate infarctions occurred in 27 patients with scans classified as IC-, and 5 patients assigned to the NP group. This is not unexpected given that the ML algorithm is derived from a system that was initially developed to detect acute intracranial hemorrhage. In addition, early ischemia may not create CT scan abnormalities that are detectable by even experienced neuroradiologists, even though ischemic stroke may produce severe neurological deficits [21, 22]. It is thus possible that cases with early acute ischemia did not result in sufficient CT abnormalities that could be detected by the ML algorithm. However, progress has been made to train an algorithm specifically to detect subtle CT changes from early cerebral ischemia. Our group has created and validated such an algorithm [22]. We foresee aggregating these algorithms to provide a more thorough assessment of patients who present with symptoms that may suggest a stroke. Such a suite of algorithms would be especially valuable in settings where a neuroradiologist may not be immediately available.

Potential clinical applications

The head CT ML algorithm described here performs sufficiently well with the potential to function as a radiology assistant. It is a tool that can help radiologists manage busy imaging centers such as the emergency department. It might be extremely valuable when a radiologist is not available on-site such as in an outpatient imaging center. Rapid evaluation of studies in urgent care environments is needed to properly manage patients, especially those with intracranial pathology. It is easily feasible for algorithm analyses to be available at the time that the scans appear on the image review station for radiologists and be marked as IC+, IC- or NP. Since IC+ scans have a >90% probability of intracranial hemorrhage or other urgent intracranial pathology as well as one-third of the NP scans, the radiologist’s attention can immediately focus on the most impactful scans. The radiologist would also know that the IC- scans are usually normal, but that small hemorrhages may be present. This may help the radiologist to adjust the search pattern to look for more subtle bleeds.”

Other ML algorithms have been shown to reduce the turnaround time for the identification and interpretation of intracranial hemorrhage or other urgent intracranial abnormalities such as intracranial hemorrhage on head CTs [23]. However, many algorithms have the limitation of only prioritizing studies with a single intracranial abnormality such as hemorrhage, which can lead to inadvertent delays in the interpretation of studies with other intracranial abnormalities that require urgent treatment. For instance, in a clinical practice where an ICH detection algorithm is implemented for worklist prioritization, a study with a large non-hemorrhagic mass with herniation requiring immediate attention may be shifted to the lower end of the reading queue due to the absence of ICH. Our ML algorithm is more likely to expedite the identification of patients with more urgent needs for admission and intervention. This system also allows flexibility in additional types of clinical settings. For instance, practices that have a longer time interval between the study scan time to radiologist interpretation, such as outpatient imaging centers, may triage IC+ and NP cases to ensure more timely identification of an intracranial hemorrhage or other urgent intracranial abnormality, which could otherwise wait for some time on a radiology worklist before the images are seen by anyone. This algorithm would be especially valuable in emergency and urgent care centers where a radiologist may not be available when the patient is scanned.

Shortcomings and future directions

The algorithm has shortcomings. Most importantly the algorithm was trained only on scans where hemorrhage was present or absent. Despite the absence of training using scans with other pathologies, the algorithm performance was surprisingly good but can be improved. The obvious path forward is to retrain the algorithm using scans with other pathologies, which we are doing. For instance, one strategy would be to retrain the algorithm using more uniform datasets representing each of various pathologies The decision rules used to define classification (i.e., IC+ are those with a predicted probability of at least 0.9 on 3 slices) were based on an empirical evaluation of a test data set, which was conducted with a limited number of cases. More detailed assessment of different threshold parameters and its effect of positive and negative predictive values using a larger data set may improve the performance of the algorithm; we will investigate this further after retraining the algorithm.

To reduce selection bias, 1000 consecutive cases that were processed by the algorithm were included in the study. While most of the studies were from the emergency department, a subset of cases included inpatient, intraoperative, and outpatient cases. However, these were also deemed urgent or emergent and were directed to the emergency neuroradiology service for interpretation. The algorithm was also not used to assist in the real-time evaluation of imaging studies, so it remains uncertain how much of a clinical impact the algorithm would have in actual clinical practice. Finally, more studies are needed to determine the value of outputting any additional ML algorithm data such including predicted probability and prediction map for radiologists.

Conclusion

This ML algorithm that incorporates uncertainty can categorize head CT scans with or without intracranial hemorrhage or other urgent intracranial abnormalities with high predictive values and result in clinically relevant classification. It may help expedite patient triage and appropriate management in various clinical settings.

Citation: Yoon BC, Pomerantz SR, Mercaldo ND, Goyal S, L’Italien EM, Lev MH, et al. (2023) Incorporating algorithmic uncertainty into a clinical machine deep learning algorithm for urgent head CTs. PLoS ONE 18(3): e0281900. https://doi.org/10.1371/journal.pone.0281900

About the Authors:

Byung C. Yoon

Roles: Data curation, Formal analysis, Investigation, Writing – original draft, Writing – review & editing

Current address: Department of Radiology/Neuroradiology, Stanford/VA Palo Alto Health Care, Stanford Medical School, Stanford, CA, United States of America

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

ORICD: https://orcid.org/0000-0003-2929-5677

Stuart R. Pomerantz

Roles: Investigation, Writing – review & editing

Affiliations: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America, Mass General Brigham Data Science Office, Boston, MA, United States of America

Nathaniel D. Mercaldo

Roles: Formal analysis, Writing – original draft, Writing – review & editing

Affiliation: Massachusetts General Hospital Institute for Technology Assessment, Boston, MA, United States of America

Swati Goyal

Roles: Data curation, Methodology, Writing – review & editing

Affiliations: Mass General Brigham Data Science Office, Boston, MA, United States of America, Department of Radiology/ Information Systems, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Eric M. L’Italien

Roles: Data curation, Methodology, Writing – review & editing

Affiliations: Mass General Brigham Data Science Office, Boston, MA, United States of America, Department of Radiology/ Information Systems, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Michael H. Lev

Roles: Conceptualization, Investigation, Writing – review & editing

Affiliation: Emergency Radiology & Neuroradiology Divisions, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Karen A. Buch

Roles: Investigation, Writing – review & editing

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Bradley R. Buchbinder

Roles: Data curation, Investigation, Writing – review & editing

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

John W. Chen

Roles: Data curation, Investigation, Writing – review & editing

Affiliations: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America, Massachusetts General Hospital Center for Systems Biology (CSB), Boston, MA, United States of America

John Conklin

Roles: Data curation, Investigation, Writing – review & editing

Affiliation: Emergency Radiology & Neuroradiology Divisions, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Rajiv Gupta

Roles: Investigation, Writing – review & editing

Affiliations: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America, Massachusetts General Hospital Consortia for Integration of Medicine and Innovative Technologies (CIMIT), Boston, MA, United States of America, Massachusetts General Hospital CT Innovation and Advanced X-ray Imaging Science (AXIS) Center, Boston, MA, United States of America

George J. Hunter

Roles: Investigation, Writing – review & editing

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Shahmir C. Kamalian

Roles: Investigation, Writing – review & editing

Affiliation: Emergency Radiology & Neuroradiology Divisions, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Hillary R. Kelly

Roles: Investigation, Writing – review & editing

Affiliations: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America, Department of Radiology, Massachusetts Eye and Ear Institute, Harvard Medical School, Boston, MA, United States of America

Otto Rapalino

Roles: Investigation, Writing – review & editing

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

ORICD: https://orcid.org/0000-0002-5038-4183

Sandra P. Rincon

Roles: Investigation, Writing – review & editing

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Javier M. Romero

Roles: Investigation, Writing – review & editing

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Julian He

Roles: Data curation, Investigation

Affiliation: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America

Pamela W. Schaefer

Roles: Investigation, Writing – review & editing

Affiliations: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America, Mass General Brigham Enterprise Radiology, Boston, MA, United States of America

Synho Do

Roles: Conceptualization, Investigation, Methodology, Writing – review & editing

Affiliation: Mass General Brigham Data Science Office, Boston, MA, United States of America

Ramon Gilberto González

Roles: Conceptualization, Formal analysis, Methodology, Project administration, Supervision, Writing – original draft, Writing – review & editing

E-mail: [email protected]

Affiliations: Neuroradiology Division, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States of America, Mass General Brigham Data Science Office, Boston, MA, United States of America, Massachusetts General Hospital Athinoula A. Martinos Center for Biomedical Imaging, Boston, MA, United States of America

ORICD: https://orcid.org/0000-0002-0234-3897

References

1. Zhou Z, Sanders JW, Johnson JM, Gule-Monroe MK, Chen MM, Briere TM, et al. Computer-aided Detection of Brain Metastases in T1-weighted MRI for Stereotactic Radiosurgery Using Deep Learning Single-Shot Detectors. Radiology. 2020;295(2):407–15. pmid:32181729

2. Stib MT, Vasquez J, Dong MP, Kim YH, Subzwari SS, Triedman HJ, et al. Detecting Large Vessel Occlusion at Multiphase CT Angiography by Using a Deep Convolutional Neural Network. Radiology. 2020;297(3):640–9. pmid:32990513

3. Arbabshirani MR, Fornwalt BK, Mongelluzzo GJ, Suever JD, Geise BD, Patel AA, et al. Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration. NPJ Digit Med. 2018;1:9. pmid:31304294

4. Rava RA, Seymour SE, LaQue ME, Peterson BA, Snyder KV, Mokin M, et al. Assessment of an Artificial Intelligence Algorithm for Detection of Intracranial Hemorrhage. World Neurosurg. 2021;150:e209–e17. pmid:33684578

5. Seyam M, Weikert T, Sauter A, Brehm A, Psychogios MN, Blackham KA. Utilization of Artificial Intelligence-based Intracranial Hemorrhage Detection on Emergent Noncontrast CT Images in Clinical Workflow. Radiol Artif Intell. 2022;4(2):e210168. pmid:35391777

6. Heit JJ, Coelho H, Lima FO, Granja M, Aghaebrahim A, Hanel R, et al. Automated Cerebral Hemorrhage Detection Using RAPID. American Journal of Neuroradiology. 2021;42(2):273–8. pmid:33361378

7. Campagner A, Cabitza F, Ciucci D. Three-Way Decision for Handling Uncertainty in Machine Learning: A Narrative Review. Rough Sets: Springer International Publishing; 2020. p. 137–52.

8. Hüllermeier E, Waegeman W. Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods. Machine Learning. 2021;110(3):457–506.

9. Kompa B, Snoek J, Beam AL. Second opinion needed: communicating uncertainty in medical machine learning. npj Digital Medicine. 2021;4(1). pmid:33402680

10. Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition2014 September 01, 2014:[arXiv:1409.556 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2014arXiv1409.1556S.

11. He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016. p. 770–8.

12. Szegedy C, Ioffe S, Vanhoucke V, Alemi A. Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning2016 February 01, 2016:[arXiv:1602.07261 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2016arXiv160207261S.

13. Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the Inception Architecture for Computer Vision. 2016:2818–26.

14. Lee H, Yune S, Mansouri M, Kim M, Tajmir SH, Guerrier CE, et al. An explainable deep-learning algorithm for the detection of acute intracranial haemorrhage from small datasets. Nature Biomedical Engineering. 2019;3(3):173–82. pmid:30948806

15. Lin M, Chen Q, Yan S. Network In Network2013 December 01, 2013:[arXiv:1312.4400 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2013arXiv1312.4400L.

16. Nesterov Y. A method for unconstrained convex minimization problem with the rate of convergence o(1/k^2). Doklady AN USSR. 1983;269:543–7.

17. Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, et al. TensorFlow: A system for large-scale machine learning 2016 May 01, 2016:[arXiv:1605.08695 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2016arXiv160508695A.

18. Wang X, Peng Y, Lu L, Lu Z, Bagheri M, Summers RM. ChestX-Ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. 2017. p. 3462–71.

19. Sozykin K, Protasov S, Khan A, Hussain R, Lee J. Multi-label Class-imbalanced Action Recognition in Hockey Videos via 3D Convolutional Neural Networks2017 September 01, 2017:[arXiv:1709.01421 p.]. Available from: https://ui.adsabs.harvard.edu/abs/2017arXiv170901421S.

20. Højsgaard S, Jun Y, Ulrich H. The R package GEEPACK for generalized estimating equations. Journal of Statistical Software. 2005;15.

21. Shen J, Li X, Li Y, Wu B. Comparative accuracy of CT perfusion in diagnosing acute ischemic stroke: A systematic review of 27 trials. PLoS One. 2017;12(5):e0176622. pmid:28520753

22. Gauriau R, Bizzo BC, Comeau DS, Hillis JM, Bridge CP, Chin JK, et al. Head CT deep learning model is highly accurate for early infarct estimation. Sci Rep. 2023;13(1):189. pmid:36604467

23. O’Neill TJ, Xi Y, Stehel E, Browning T, Ng YS, Baker C, et al. Active Reprioritization of the Reading Worklist Using Artificial Intelligence Has a Beneficial Effect on the Turnaround Time for Interpretation of Head CT with Intracranial Hemorrhage. Radiol Artif Intell. 2021;3(2):e200024. pmid:33937858