Quantized ML Models for MRI Brain Tumor Classification in Low-Resource Settings

A paper available on arXiv (2605.19207) introduces a comprehensive compression framework aimed at implementing deep learning models in resource-limited clinical settings. This framework focuses on classifying brain tumors using MRI scans and incorporates techniques such as quantization-aware training, knowledge distillation from a DenseNet-101 model to a DenseNet-32 model with low-bit post-training quantization, along with Float16 post-training quantization utilizing a MobileNetV2 backbone. The research utilizes a multi-class MRI dataset that includes glioma, meningioma, pituitary tumors, and healthy controls. Extensive validation of the MobileNetV2 pipeline is presented, which employs a three-stage transfer learning approach and implements Float16 quantization through TensorFlow Lite, while also detailing the DenseNet-based strategies. This study addresses the computational, memory, and power limitations that restrict the application of deep learning models in low-resource healthcare environments.