Medfusion-Net: A Dual-Path CNN-Transformer Model with Multi-Scale Attention for Medical Image Analysis
H. Srimathi
Abstract
Purpose: Medical image analysis increasingly relies on deep learning for classification and decision-support tasks, but models must balance local texture recognition, global contextual reasoning, computational efficiency, and interpretability. This paper proposes MedFusion-Net, a new dual-path deep learning architecture designed for medical image classification. The model is intended to combine convolutional feature extraction with transformer-based global context modelling and multi-scale attention so that clinically relevant local and global patterns can be represented jointly. Design/Methodology/Approach: The proposed framework uses a CNN branch for hierarchical local feature extraction and a lightweight transformer branch for long-range dependencies. Multi-scale channel-spatial attention is introduced before feature fusion, followed by a gated fusion module and classification head. The proposed model is designed for reproducible evaluation using publicly available medical-image benchmarks such as MedMNIST v2, with patient-level or officially supplied train-validation-test splits preserved where applicable. Evaluation is specified using accuracy, precision, recall, F1-score, ROC-AUC, parameter count, inference cost, and calibration/error analysis. Baselines include conventional CNNs and representative CNN-transformer architectures. Findings: This manuscript presents the architecture, training protocol, evaluation design, and analytical framework for validating MedFusion-Net. No numerical experimental results are fabricated in this manuscript. The proposed contribution is therefore a model and a reproducible validation protocol; numerical claims should be inserted only after the experiments have been executed on the selected dataset(s). The design is motivated by evidence that transformer-based approaches can capture long-range relationships that are difficult for purely local convolutional operators, while CNNs remain effective for fine-grained spatial patterns. Practical Implications: The framework can support researchers developing lightweight medical-image decision-support systems and can serve as a baseline for investigating the trade-off between accuracy, computational cost, and explainability. The proposed attention and visualization components can also be used to inspect whether predictions are driven by clinically meaningful image regions. Originality/Value: The paper proposes a unified dual-path architecture in which local convolutional representations and global transformer representations are selectively fused using multi-scale channel-spatial attention and a gated fusion mechanism. The principal novelty is the architectural integration and the explicit evaluation of accuracy, efficiency, calibration, and explanation quality rather than accuracy alone.