Experimental Evaluation of Deep Multi-Scale Fusion Networks for Early Detection of Tumor Lesions based on Medical Images

The detection of tumor lesions at an early and accurate point is a decisive factor in enhancing treatment outcomes in cancer therapy. This paper introduces a new Deep Multi-Scale Fusion Network (DMFN) that is capable of maximizing the performance of tumor classification by using fine-grained and global contexts of medical images. Based on the BraTS dataset, the suggested approach takes out lesioncentered patches across several spatial scales (128 × 128) and processes them with parallel CNN branches. The scalespecific features are combined with the help of an attention mechanism, making the model adjustment-based on attentively prioritizing the most informative representation of different tumor types. Extensive experiments with baselines like classifiers and deep networks like ResNet-50, U-Net, and EfficientNet show that DMFN performs better. It obtains a complete classification accuracy of 96.5, and a precision of 96.1, recall of 95.9, and F1-score of 96.0. Ablation experiments are used to prove the role of multiscale fusion and attention layers in enhancing predictive accuracy. Moreover, the fact that lesion localization is done using heatmap also contributes to the model interpretability.