Explainable Dual-Feature Knowledge Distillation for Efficient Autism Spectrum Disorder Classification Using fMRI Data

Fatima Ez-Zahraa Bazay, Houda El Mohamadi, Mohammed El Hassouni, Rachid Jennane, Ahmed Drissi El Maliani

Keywords: Knowledge Distillation (KD), Explain ability, Saliency maps, Texture Shape features, fMRI data, ASD

Abstract:

Deep learning models for Autism Spectrum Disorder (ASD) classification from functional Magnetic Resonance Imaging (fMRI) data face two critical barriers to clinical deployment: high computational costs and lack of interpretability. We propose Dual-Feature Knowledge Distillation (DFKD), a framework that transfers both predictive accuracy and explainability from large teacher models to compact student
networks. DFKD leverages Dual-Feature Saliency Extraction (DFS-Ex) to capture complementary texture and shape features from fMRI-derived glass brain visualizations, guiding student training through spatial attention alignment. Evaluation on ABIDE across eight teacher-student pairs demonstrates that DFKD achieves up to 97.95% accuracy with 8.6× compression, notably improving ResNet101-GhostNet from 93.46% baseline to 97.95%, consistently outperforming conventional distillation methods (Kullback-Leibler divergence Knowledge Distillation (KL-Div), Intermediate Knowledge Distillation (I-KD)). Grad-CAM visualizations confirm DFKD-trained students inherit interpretable attention patterns, focusing on diagnostically relevant brain regions. Our approach enables deployment of efficient, transparent
models in resource-constrained clinical environments.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Interpretability and Explainable AI

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 384