Authors:
Abhishek Mishra, N. Sankar Ram, Mahin Raju, S. Sham Karthik, Raad Shahamat Chowdhury, Varun Kumar Nomula
Addresses:
Department of Computer Science and Engineering in Artificial Intelligence and Machine Learning, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. Department of Software Engineering, Victoria University, Sydney, New South Wales, Australia. Department of Analytics/Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, Georgia, United States of America.
Abstract:
GPU dependence makes deep learning dermoscopy models fail in resource-limited clinical settings such as mobile screening units, rural health posts, and primary-care clinics. The CPU-only Compression-Driven Pipeline Framework (CDPF) segmentation-classification architecture fills this deployment gap for dermoscopic image processing. A 97.3-million-parameter hybrid backbone with U-Net++ segmentation and an MGSAF classification head reduces GPU dependency. Structured L1-norm channel pruning reduces FLOPs from 97.3M to 68.1M and 18.4G to 12.9G. KL-divergence and mean squared error on intermediate feature tensors address accuracy loss in feature-level knowledge distillation. The 4x memory bandwidth reduction of INT8 arithmetic improves CPU latency in INT8 static quantisation with ONNX Runtime export. The 61.4-million-parameter model on an Intel Core i7-12700 CPU achieves 89.3% Dice Similarity Coefficient (DSC) and 85.0% macro-averaged F1-score in HAM10000 (10,015 images, 7 classes) and ISIC 2019 (25,331 images, 9 classes) without GPU acceleration, with a 7.9× latency drop over the uncompressed FP32 baseline. Ablation studies show that pruning reduces latency by 29.9%, distillation improves accuracy at no cost, and quantisation speeds up inference by 7.9×. Grad-CAM saliency maps validated by a dermatologist match clinically recognised dermoscopic features in 92% of randomly selected test cases, demonstrating the potential of compression-first design for research-grade AI performance and limited clinical deployment.
Keywords: Knowledge Distillation; Limited Clinical Deployment; Ablation Studies; Quantisation Speeds; Dice Similarity Coefficient (DSC); Compression-First Design.
Received on: 02/09/2025, Revised on: 21/10/2025, Accepted on: 04/11/2025, Published on: 07/06/2026
DOI: 10.64091/ATICR.2026.000342
AVE Trends in Intelligent Computing Research, 2026 Vol. 1 No. 2 , Pages: 102-113