AMS-ETL: An Adaptive Multi-Source Ensemble Transfer Learning Framework for Robust Multi-Disease Diagnostic Classification

Kishore Babu Nampalle, Dhiran Kumar Mahto

Keywords: Adaptive Ensemble Learning, Transfer Learning, Medical Image Classification, Multi-Disease Diagnosis, Convolutional Neural Networks (CNN), Robust Classification

TL;DR: A transfer learning–based deep ensemble model for medical image classification

Abstract: Deep learning methods for medical image analysis, while powerful, often face performance limitations due to dataset constraints, including small size, quality variability, and heterogeneous backgrounds. This is particularly critical in diagnosing severe conditions like skin cancer and brain tumors, where predictive reliability is paramount. Current approaches leveraging single-model transfer learning can struggle with generalization, while conventional ensembles often aggregate pre-trained models without optimizing their complementary strengths for specific clinical imaging characteristics. To address this, we propose Adaptive Multi-Source Ensemble Transfer Learning (AMS-ETL), a framework that strategically integrates diverse pre-trained architectures and employs a meta-learning strategy for dynamic source-weighting. Our method uniquely tailors the ensemble composition by evaluating each model’s discriminative capacity for specific feature patterns in clinical images, moving beyond simple performance averaging. We implement this using a foundational MobileNet feature extractor combined with auxiliary sources, processed through a gated logistic regression meta-learner for final prediction. Validation on clinical dermatology and neuroimaging datasets demonstrates that AMS-ETL achieves state-of-the-art accuracy and significantly improves robustness against overfitting. Furthermore, our model provides enhanced feature diversity and discriminative interpretability, offering clinicians granular decision support. This work establishes that adaptive, source-aware ensemble design is crucial for advancing automated, reliable diagnostic frameworks in resource-constrained clinical environments.

Primary Subject Area: Transfer Learning and Domain Adaptation

Secondary Subject Area: Learning with Noisy Labels and Limited Data

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 312