End-to-end machine learning based discrimination of neoplastic and non-neoplastic intracerebral hemorrhage on computed tomography

Abstract

Purpose

To develop and evaluate a fully automated segmentation and classification tool for the discrimination of neoplastic and non-neoplastic intracerebral hemorrhage (ICH) on admission Computed Tomography (CT).

Materials and methods

Two models were developed using a retrospective dataset of acute ICH patients with unknown etiology upon admission, based on CT scans from a single institution (January 2016 to May 2020). An nnU-Net segmentation model was trained on manually segmented ICH and perihematomal edema (PHE) masks, alongside a ResNet-34 classification model for differentiating between neoplastic and non-neoplastic ICH. The combined tool was evaluated on the test set and validated on an external cohort. Validation performance was reevaluated after enriching the training data of the segmentation model. Evaluation metrics included accuracy (Acc), area under the curve (AUC), sensitivity, specificity, and Matthews Correlation Coefficient (MCC). Performance was compared to human raters.

Results

Among 291 patients, 116 (39.86 %) had neoplastic and 175 (60.14 %) non-neoplastic ICH. The 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 validation cohort (n = 58), the tool achieved an AUC of 68 % reaching 83 % after retraining of the segmentation model. The tool achieved an MCC of 0.62, compared to 0.47–0.61 for the human raters.

Conclusion

The tool demonstrated high diagnostic performance with potential as a decision-aiding tool; however, it relies on multi-vendor data for improved robustness, warranting further validation across diverse datasets.