Beyond Prediction: An Agentic Adaptive CNN for Skin Cancer Diagnosis and Autonomous Clinical Workflow Initiation

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Skin cancer incidence continues to rise worldwide, and delayed diagnosis remains a principal determinant of mortality. This paper proposes a detection system based on an adaptive Convolutional Neural Network (ACNN) that classifies skin cancer from lesion images using gradient based optimization. The framework is organized according to agentic AI principles, structured as a loop spanning four stages, namely perception, decision, action, and feedback, rather than a single forward prediction. Lesion images are perceived and passed through forward propagation, where a fitness function governs weight adaptation, continuously regenerating updated weights instead of relying on a fixed trained state. The decision stage yields probabilistic benign and malignant classifications, which are routed to a multi stage output panel that autonomously executes the corresponding clinical action: triggering specialist referral, scheduling biopsy, dispatching the case to a dermoscopy bot, or notifying the patient. The system therefore advances the diagnostic pathway rather than terminating in a label. Feedback closes the loop, as classification accuracy is monitored against a defined threshold, and any shortfall reinvokes the backward pass, re deriving weights through the fitness function and returning them to the forward pass for a further cycle. This threshold driven self correction allows the model to evaluate and revise its own performance without external supervision, reflecting the autonomy and goal directed behaviour characteristic of agentic AI. Simulation results on the HAM10000 dataset show that the proposed ACNN attains higher accuracy and lower training loss than baseline CNN and SVM models, providing substantive decision support to clinicians.