Lightweight Deep Learning Models for Brain Tumor Classification

Differentiating brain tumours by MRI using computer algorithms remains a huge challenge in clinical neuro- oncology and medical imaging area. In our work, we proposed a new ultra-lightweight convolutional neural network (CNN) deep structure for practical use in clinical setting. Our CNN embeddings three private self- intelligent modules: (1) a Micro Adaptive Feature Extractor applied spatial-channel attention to whole image to have a dynamic feature refinement; (2) a Micro Multi-Scale Processor that replicates expert radiologist diagnostic manoeuvres by enabling learned multi-resolution feature combination; and (3) a Context-Aware Intelligent Pooling that offers content-based flexible pooling strategies. The complete Mendeley Brain Tumour MRI dataset (n=12,064 images) was used to test each approach, and resulted with an excellent classification accuracy of 99.22%. Such Mendeley Brain Tumour MRI dataset had an extremely tiny footprints of only 659,228 trainable parameters, which was 80% less intricate than the other architectures such as U-Net. The model demonstrate that it could be effective on three independent external validation datasets Kaggle (99.24%), Figshare (87.47%) and Br35H (98.33%). This demonstrates that the model can be applied across datasets. We achieved several essential features needed for clinical deployment for reducing inference latency (12ms), a smaller model footprint (2.6 MB), Achieved good results given the system complexity (network size, operation count, parameter size) and the frugal memory utilization (4.4 MB RAM). Ablation analysis proved that the intelligent pooling module contributed to accuracy increase up to 2.18%. This study represents a significant step forward in parameters-efficient deep learning for medical imaging, enabling feasible deployment on mobile computing platforms, edge computing devices, and in norm constrained healthcare centres around the world.