Brain Tumor Identification Using Capsule Networks in Medical Imaging Diagnostics
Annotatsiya
Convolutional neural networks, the dominant architecture for medical image classification, discard fine-grained spatial pose information through max-pooling operations-a design choice that improves translation invariance but may discard the part-whole and orientation relationships that carry diagnostic information in tumor morphology. Capsule Networks (CapsNet), introduced by Sabour, Frosst, and Hinton (2017), replace max-pooling with dynamic routing-by-agreement between capsules-vector-valued units whose length encodes entity-existence probability and whose orientation encodes pose-offering a representational alternative explicitly designed to preserve such relationships. This paper applies CapsNet to four-class brain tumor identification (Glioma, Meningioma, Pituitary Tumor, No Tumor) on a multi-institutional T1-weighted contrast-enhanced MRI dataset of 10,000 annotated slices collected from six radiology departments over 24 months (Glioma: 2,800; Meningioma: 2,200; Pituitary: 1,800; No Tumor: 3,200), with three dynamic routing iterations, margin loss, and a reconstruction-regularised decoder. On a held-out test set of 2,000 slices, CapsNet achieves 91.2% accuracy, 89.7% sensitivity, 92.4% specificity, and$\mathbf{A U C}=\mathbf{0. 9 4 4}$, outperforming ResNet-50 (88.4% accuracy; +3.2% relative), VGG-16 (84.8%; +7.5%), and a parameter-matched standard CNN (82.4%; +10.7%) while using only 8.2 million parameters-3.1 times fewer than ResNet-50's 25.6 million. A radiologist-in-the-loop validation study found that pairing CapsNet's output with senior radiologists improved their standalone accuracy from 92.4% to 96.8% (+4.8% relative) while reducing read time by 33.3%, with a substantially larger benefit for junior radiologists (84.6% to 94.4%, +11.6% relative; read time reduced 47.1%), consistent with the broader medical AI literature's finding that decision support narrows experience-related performance gaps more than it widens already-strong performance. Failure case analysis of the 176 misclassified test slices, cross-scanner generalisation testing, and an out-of-taxonomy robustness check on rare tumor subtypes not represented in training each surface distinct boundaries on these results, the last of which identifies a high-confidence misclassification failure mode requiring an explicit safeguard before any clinical deployment.
Hali tarjima qilinmagan