TLDR: This paper presents a deep learning solution for identifying morpho-pathological features in glioblastoma, developed for the BraTS-Path Challenge 2024. The researchers used a fine-tuned ResNet-18 model on a reclassified dataset of H&E-stained tissue sections. While the model showed high performance on a local validation set, its accuracy, recall, and F1-score were lower on the challenging Synapse online platform, though it maintained high specificity and secured second place in the challenge. The study highlights the potential of deep learning in glioblastoma diagnosis but also points to the need for models with better generalization across diverse data sources.
Glioblastoma, an aggressive and complex brain tumor, presents significant challenges in accurate diagnosis and treatment due to its diverse molecular and pathological features. Traditional diagnostic methods often struggle to fully capture this complexity, making it difficult to choose the most effective therapies and improve patient outcomes. However, deep learning offers a promising avenue for enhancing glioblastoma diagnosis.
In response to this challenge, the BraTS-Pathology Challenge 2024 was launched to encourage the development of AI tools capable of identifying distinct histological regions within brain tumors. A team from the University of Nottingham Ningbo China, comprising Juexin Zhang, Ying Weng, and Ke Chen, participated in this challenge, presenting their solution focused on leveraging deep learning for glioblastoma morpho-pathological feature identification. You can find their full paper here: Deep Learning for Glioblastoma Morpho-pathological Features Identification.
The Approach: Transfer Learning with ResNet-18
The researchers adopted a transfer learning approach, utilizing a pre-trained ResNet-18 model. This model, initially trained on the vast ImageNet dataset, was then fine-tuned using the BraTS-Path training dataset. The core of their model architecture involves multiple residual blocks, a hallmark of ResNet design, which helps in processing complex image features. The model was designed to classify specific histological areas of interest within glioblastoma tissue samples.
Dataset and Key Features
The BraTS-Path challenge dataset consists of H&E-stained FFPE digitized tissue sections from The Cancer Imaging Archive’s TCGA-GBM and TCGA-LGG collections. These collections were reclassified to align with the latest 2021 World Health Organization (WHO) classification of central nervous system tumors, focusing on glioblastoma multiforme (GBM) cases lacking the IDH mutation (IDH-wildtype) and specific low-grade astrocytomas. Expert neuropathologists meticulously annotated these tissue sections, dividing them into uniform patches categorized by specific tissue types or labeled as ‘background’. The six key annotated histological areas of interest include: cellular tumor (CT), pseudopalisading necrosis (PN), microvascular proliferation (MP), geographic necrosis (NC), infiltration into the cortex (IC), and penetration into white matter (WM).
Before feeding the images to the model, a series of preprocessing steps were applied. This included converting image files from BGR to RGB color space, scaling pixel intensity values to a range, and applying dataset normalization to center pixel values at zero mean and scale them to unit standard deviation. These steps are crucial for numerical stability and enhancing model generalization.
Performance and Results
The model’s performance was evaluated using several metrics, including accuracy, recall, F1-score, specificity, and the Matthews Correlation Coefficient (MCC). On their local validation set, the model showed promising results with high accuracy, recall, and F1-score of 0.9872, a specificity of 0.9974, and an MCC of 0.9828. However, when assessed on the more challenging Synapse online validation platform, the performance dropped. The model achieved an accuracy, recall, and F1-score of 0.392229, a specificity of 0.898704, and an MCC of 0.255267. Despite this drop, their solution notably secured second place during the testing phase of the BraTS-Path Challenge.
Also Read:
- Systematic Review Explores Multimodal Machine Learning for Cancer Survival Prediction
- SmartPath-R1: A New AI System for Comprehensive Pathology Analysis
Conclusion and Future Directions
While the model demonstrated strong performance on local validation data, its results on the online validation set suggest limitations in generalizing to data from various sources. This highlights the inherent difficulty in diagnosing glioblastoma due to its aggressive nature and complex heterogeneity. The study underscores the potential of deep learning in enhancing glioblastoma diagnosis and emphasizes the need for future research to focus on developing models capable of consistent performance across diverse datasets to improve reliability and clinical applicability.
