End-to-End Machine Learning based Discrimination of Neoplastic and Non-neoplastic Intracerebral Hemorrhage on Computed Tomography
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Purpose
To develop and evaluate an automated segmentation and classification tool for the discrimination of neoplastic and non-neoplastic intracerebral hemorrhage (ICH) on admission Computed Tomography (CT) utilizing images containing hemorrhage and perihematomal edema.
Materials and Methods
The models were developed and evaluated using a retrospective dataset of patients who presented with acute ICH of unknown cause upon admission, using CT scans obtained from a single institution between January 2016 and May 2020 for both training and testing. Etiology of ICH were binarized into non-neoplastic and neoplastic ICH according to follow-up MRI results based on the ATOMIC ICH classification. Masks for ICH and PHE were manually segmented. Two separate models were trained: 1) An nnU-Net segmentation model 2) A ResNet-34 classification model. An end-to-end tool was evaluated by concatenating the two models which allowed the segmentation model to preprocess the images for the classification model. Performance enhancement was assessed by fine-tuning the model on a randomly selected, small subset of the external cohort. To assess the model’s generalizability, the performance was additionally validated on an external dataset. Evaluation metrics were accuracy (Acc), area under the curve (AUC) and corresponding sensitivities and specificities.
Results
A total of 291 patients were included of whom 116 (39.86%) presented with neoplastic and 175 (60.14%) with non-neoplastic ICH. The end-to-end classification tool achieved an Acc of 86% and an AUC of 85% with a sensitivity and specificity of 80% and 93% in the test set. On the external validation cohort (n=58), the classification pipeline achieved an AUC of 68% and Acc of 66% (sensitivity 64%; specificity 67%). Fine-tuning on a selected small subset of the external cohort enhanced performance, achieving an AUC and accuracy of 70% (sensitivity 70%; specificity 71%).
Conclusion
An end-to-end classification tool achieved a high diagnostic performance and generalizability in classifying neoplastic from non-neoplastic ICH on CT, suggesting a robust framework for a potential clinical implementation as a decision-aided tool in early ICH management.
