Diffuse gliomas, comprising astrocytoma, oligodendroglioma, and glioblastoma, represent a significant proportion of malignant brain tumors in adults. The prognosis varies across these subtypes, with median survival ranging from 60 to 119 months in oligodendroglioma, 18 to 36 months in astrocytoma, and a stark 8 months in glioblastoma.
Challenges in Current Diagnostic Approaches:
The diagnostic process for diffuse gliomas involves a meticulous examination of both histological and molecular features, making it a time-consuming, labor-intensive, and economically burdensome procedure for patients. Histological diagnosis, reliant on pathologists’ expertise, is prone to interobserver variation. Molecular diagnosis, on the other hand, often requires invasive surgical procedures for tissue extraction, followed by complex sequencing techniques that may not be readily available in all medical centers.
Digital Transformation and Computational Pathology:
The advent of digitized scanners has paved the way for the transformation of glass slides into whole-slide images (WSIs), presenting an opportunity for automated computational pathology. Existing models typically utilize convolutional neural networks (CNNs) for image recognition. While various CNN models for glioma diagnosis have been proposed, a comprehensive WSI diagnostic model strictly adhering to the 2021 WHO classification is still lacking.
Unique Challenges and Solutions:
WSIs pose unique challenges due to their gigapixel-level resolution, making traditional CNNs computationally unfeasible. To overcome this, WSIs are divided into smaller patches, and weakly supervised learning techniques are employed to train CNNs using coarse labels such as cancer or non-cancer. This methodology helps avoid the manual annotation burden by leveraging patient-level tumor types as weak supervision labels.
Proposed Neuropathologist-Level Integrated Diagnosis Model:
In this groundbreaking study, a neuropathologist-level integrated diagnosis model is introduced for the automated prediction of 2021 WHO types and grades of adult-type diffuse gliomas from annotation-free standard WSIs. The model sidesteps the annotation burden by using patient-level tumor types as weak supervision labels and capitalizes on type-discriminative patterns through feature domain clustering.
Validation and Results:
The integrated diagnosis model is rigorously developed and externally tested using a dataset comprising 2624 patients with adult-type diffuse gliomas from three hospitals. All datasets adhere strictly to the 2021 WHO classification requirements, encompassing integrated histopathological and molecular information. The results showcase the potential of the proposed model as an automated and unbiased classification tool for adult-type diffuse gliomas.
Discussion:
Our study introduces a novel CNN-based integrated diagnosis model for adult-type diffuse gliomas, aligned with the 2021 WHO classification. The strengths of our research lie in its ability to overcome key challenges in computational pathology, presenting several advantages compared to previous works.
Firstly, our model leverages a patch clustering technique, allowing training with only tumor types as weakly supervised labels. This innovative approach eliminates the need for pixel-level or patch-level annotations, reducing the annotation burden and enhancing the model’s scalability.
Secondly, the model achieves high-performance integrated diagnosis solely from pathological images, eliminating the necessity of combining pathological and molecular information separately. This accomplishment is attributed to the clustering-based CNN, which effectively learns imaging features encompassing both pathological morphology and underlying biological cues.
Thirdly, the extensive dataset, consisting of 644,896 patch images from 1362 patients, empowers our model to generalize well across internal and external testing cohorts. It demonstrates robust performance in classifying major types, grades within types, and notably in distinguishing genotypes with shared histological features.
The 2021 WHO classification introduces a nuanced “grades within type” system, combining histological and molecular data for classification. Remarkably, our model predicts tumor grades/types directly from pathological images, showcasing its capability to learn molecular characteristics solely from the visual representation of tissues.
The clustering-based CNN model used in our study presents significant advantages. Firstly, it eliminates the need for manual annotation, automatically selecting type-relevant patch clusters crucial for integrated classification. Secondly, by aggregating local features, the model reaches a global diagnosis, selectively fusing the most discriminative information from multiple relevant patches.
Comparative analysis with classical MIL and its variants, AMIL and CLAM, highlights the superior performance of our clustering approach in classifying integrated types, histological categories, and grades within each type. The attention mechanism incorporated in AMIL and CLAM did not yield expected benefits, possibly due to the high variability and complexity within the pathological data.
Despite these accomplishments, three limitations should be acknowledged. First, the dataset, though substantial (2624 patients), could benefit from expansion through future international multicenter and multiracial collaborations. Second, to enhance model robustness, we acknowledge the need for a larger dataset obtained from various scanners, necessitating advanced stain normalization methods. Third, further preclinical experimental work is needed to fully elucidate the biological interpretability of our deep-learning model.
In conclusion, our CNN model demonstrates the potential for fully automated integrated diagnosis of adult-type diffuse gliomas, strictly adhering to the 2021 WHO classification. The model’s scalability, generalization ability, and ability to learn molecular characteristics from pathological images make it a promising tool for unbiased classification in clinical scenarios. Future work will focus on expanding datasets, addressing scanner variability, and conducting additional preclinical experiments for a comprehensive understanding of the model’s biological interpretability.
reference link : https://www.nature.com/articles/s41467-023-41195-9#Sec8
