A Multi-Stage Deep Learning Framework for Automated Brain Tumor Diagnosis and Clinical Report Generation

Detection of brain tumors through MRI scans is a difficult and crucial process since medical imaging is complex and there are not enough experts to analyze these images in many countries. The timely detection of diseases such as gliomas, meningiomas, and pituitary tumors is critical because delayed detection may have adverse impacts on the patient’s condition and result in incorrect treatment. These difficulties are why automated deep learning models were introduced to help diagnose brain tumors. In this research, we introduce a multi-stage sequential pipeline for brain tumor classification, localization, explainability, and report generation in a radiologist-style structured format based on MRI imaging data, with each stage trained and evaluated independently on its respective dataset. Specifically, an Xception-based classifier was used to classify MRI images into one of four classes, achieving 98.96% accuracy and an F1-score of 0.9876. To enhance interpretability, the Grad-CAM technique was used to visualize image patches the model used during prediction. If the tumor was present, U-Net was used to perform localization, giving a Dice score of 0.7835 and an IoU of 0.6919 with the use of T1-weighted MRI images. Finally, the obtained features were passed as input to a QLoRA fine-tuned language model that generates structured radiology reports, including such components as Technique, Findings, and Impression.The proposed framework was evaluated on a hold-out subset of publicly available T1-weighted MRI data, achieving LLM-as-a-Judge scores of 4.92/5 for accuracy and 4.95/5 for fluency. These results suggest that the proposed approach is promising and provide an initial proof of concept. However, additional validation on independent multi-center datasets and multimodal MRI data is still needed before considering its use in clinical practice.