Abstract
Deep learning classifiers deployed in clinical settings routinely issue overconfident predictions on ambiguous or out-of-distribution inputs, posing a significant patient safety risk. This thesis proposes a post-hoc geometric abstention framework that extends the circular prototypical space approach of Dakshit (2024) by replacing the spherically symmetric boundary with a convex hull a shape-adaptive polytope that exactly encloses the true positive training embeddings of each class. The framework requires no retraining or modification of base model weights and is applicable to any pre-trained deep neural network classifier. Embeddings are extracted and projected via truncated Singular Value Decomposition into spaces of 2, 6, and 8 dimensions. A convex hull is constructed per class from correctly classified training samples, and a validation sweep identifies optimal hull expansions under nonoverlap or subtraction assignment, verified via linear programming feasibility. The framework is evaluated on three medical datasets a one-dimensional ECG rhythm dataset, a two-dimensional ECG image dataset, and a multiclass chest radiograph dataset using a custom Rhythm CNN, EfficientNetB0, and VGG16 respectively. On the ECG 1D task, the method achieves 100% retained accuracy with complete error elimination at 6D, deferring 34.9% of samples. On the chest radiograph task, the 8D projection achieves the lowest deferral rate (18.7%) and highest retained accuracy (99.71%) with 80% error improvement. On the ECG 2D task, geometric inseparability constrains the system to subtraction mode, yielding meaningful but more modest gains. These results establish convex hull prototypical spaces as a principled, interpretable, and clinically deployable foundation for selective prediction in medical AI.
Date of publication
Spring 2026
Document Type
Dissertation
Language
english
Persistent identifier
http://hdl.handle.net/10950/5051
Committee members
Dr. Yi Li , Dr. Arun Kulkarni , Dr. Sagnik Dakshit , Dr. Leonard Brown
Degree
Master of Science in Computer Science
Recommended Citation
Syed, Masroor Ul Hasan, "DEEP LEARNING ABSTENTION USING GEOMETRIC ALGORITHMS" (2026). Computer Science Theses. Paper 8.
http://hdl.handle.net/10950/5051