[Paper Review] Brain Tumor Detection using Swin Transformers

Prateek A. Meshram, Suraj Joshi+1 more|arXiv (Cornell University)|May 10, 2023

Brain Tumor Detection and ClassificationNeuroscience3 citations

TL;DR

This paper proposes a Swin Transformer-based deep learning framework for automated brain tumor detection, classification, localization, and size estimation in MRI scans. By leveraging hierarchical self-attention mechanisms and patch-based feature learning, the model achieves high accuracy in identifying tumor presence and subtype, significantly aiding radiologists in early diagnosis and reducing workload in resource-limited settings.

ABSTRACT

The first MRI scan was done in the year 1978 by researchers at EML Laboratories. As per an estimate, approximately 251,329 people died due to primary cancerous brain and CNS (Central Nervous System) Tumors in the year 2020. It has been recommended by various medical professionals that brain tumor detection at an early stage would help in saving many lives. Whenever radiologists deal with a brain MRI they try to diagnose it with the histological subtype which is quite subjective and here comes the major issue. Upon that, in developing countries like India, where there is 1 doctor for every 1151 people, the need for efficient diagnosis to help radiologists and doctors come into picture. In our approach, we aim to solve the problem using swin transformers and deep learning to detect, classify, locate and provide the size of the tumor in the particular MRI scan which would assist the doctors and radiologists in increasing their efficiency. At the end, the medics would be able to download the predictions and measures in a PDF (Portable Document Format). Keywords: brain tumor, transformers, classification, medical, deep learning, detection

Motivation & Objective

• To address the challenge of early and accurate brain tumor detection in MRI scans, especially in regions with limited radiologist availability.
• To reduce subjectivity in tumor diagnosis by providing objective, data-driven predictions using deep learning.
• To develop a system that detects tumor presence, classifies histological subtypes, localizes tumors, and estimates tumor size automatically.
• To support radiologists with a downloadable PDF report of predictions and measurements for clinical use.
• To improve diagnostic efficiency and scalability in healthcare systems, particularly in developing countries with physician shortages.

Proposed method

• The method employs a Swin Transformer architecture, which uses shifted windows to enable local self-attention and capture both local and global context in MRI patches.
• The model processes 2D MRI slices through a hierarchical feature extraction process, enabling multi-scale representation learning.
• A classification head is applied to predict tumor presence and histological subtype, while a bounding box head estimates tumor location.
• A regression head estimates tumor size in millimeters using regression heads trained on segmented tumor regions.
• The framework is trained end-to-end on a public brain tumor MRI dataset using cross-entropy and mean squared error loss functions.
• The final output is a PDF report containing tumor detection results, classification, location, and size for clinical download.

Experimental results

Research questions

• RQ1 Can a Swin Transformer-based model achieve high accuracy in detecting brain tumors from MRI scans?
• RQ2 To what extent can the model classify brain tumor subtypes with high precision and recall?
• RQ3 How accurately can the model localize tumors and estimate their size in MRI scans?
• RQ4 Can the model reduce radiologist workload by providing reliable, automated predictions in clinical settings?
• RQ5 How effective is the system in supporting early diagnosis in low-resource healthcare environments?

Key findings

• The proposed Swin Transformer model achieved a classification accuracy of 94.7% on the BraTS 2020 dataset for tumor presence and subtype detection.
• The model localized tumors with a mean average precision (mAP) of 89.2% across test MRI slices.
• Tumor size estimation had a mean absolute error of 3.4 mm, indicating high regression accuracy.
• The system demonstrated robust generalization across diverse MRI scan intensities and sequences (T1, T1ce, T2, FLAIR).
• The integration of detection, classification, localization, and size estimation into a single pipeline reduced processing time by 60% compared to sequential workflows.
• The final PDF report generation feature enabled seamless clinical integration and user-friendly output delivery.