Deep Learning-Based MRI Analysis for Brain Tumor Classification
Open Access Deposited(ii) a transparent “explained-as-you-screen” workflow that can visualize the most important image pixels for the model’s classification decisions
and (iii) an analysis of transformer limitations in the context of limited image data setting.
final-layer saliency maps from the classifier were over-laid on tumor masks generated by a pretrained U-Net segmentation model. Histograms of intersection-over-union scores revealed that salient pixels clustered predominantly at tumor margins, mirroring radiologic practice and supporting Hypothesis 1 that tumors’ boundary features dominate classification results. For the fine-tuned ResNet101 model, inference time on an A100 GPU was < 3 s for a 60-slice study. The low requirement on computational resources enables real-time triage, particularly as a rapid first-pass screening tool in the clinical workflow as the model achieved excellent results in classifying non-tumor scans. The praxis contributes by developing (i) a deep learning model for margin-aware four-class brain-MRI image classification. In particular, the model has great potential to serve as a rapid and automated first-pass screening tool in the clinical workflow
Abstract of Praxis Deep Learning-Based MRI Analysis for Brain Tumor Classification Brain tumor diagnosis demands rapid and reliable analysis of medical image scans, yet routine magnetic-resonance imaging (MRI) analysis is labor-intensive, error-prone, and constrained by shortages of expert neuroradiologists. This praxis developed an end-to-end deep-learning model that automatically classifies brain MRI slices into four categories—no-tumor, meningioma, glioma, and pituitary adenoma and additionally explains the image regions that drive each decision with saliency map analysis. Two public datasets were merged (3064 brain tumor MRI scans and 1162 healthy brain MRI scans), followed by pre-processing, data cleaning, and image augmentation with geometric and intensity transformations. A ResNet-101 backbone, pretrained on ImageNet, was fine-tuned by unfreezing the deepest residual block and attaching a four-node soft-max head. Hyper-parameter search and regularization techniques yielded a model with 99.1 % overall accuracy, macro-precision 98.9 %, and macro-recall 99.2 % on a 20 % patient-wise hold-out test set- surpassing the 95 % target. Vision-transformer (Swin) and two-stage transformer hybrids were evaluated but under-performed (< 75 % tumor-type accuracy), underscoring the efficiency of residual CNNs for modest medical datasets. Interpretability of the fine-tuned ResNet101 model was assessed with saliency map as well as through a dual-model strategy
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.