TLDR: A new deep learning model has been developed and extensively validated for automatically segmenting tumors in histopathological whole-slide images across multiple cancer types. Trained on over 20,000 images from four cancer types, the model achieved high accuracy (over 80% Dice coefficient) in external validation cohorts spanning six cancer types, including those not used in training. It demonstrated comparable performance to specialized single-cancer models and robustness across different scanners and sample preparations, though small, fragmented tumors remain a challenge. This research suggests that a single, universal model can effectively assist pathologists in diagnosing various cancers.
In the evolving landscape of precision medicine, digital pathology and artificial intelligence (AI) are becoming indispensable tools. Pathologists face increasing workloads due to rising cancer incidences and a shortage of specialists, making automated diagnostic assistance crucial. A recent study introduces a groundbreaking universal deep learning model designed to automatically identify and delineate tumor areas in histopathological whole-slide images (WSIs) across a wide range of cancer types.
Traditionally, most automatic tumor segmentation models have been developed and tested for single cancer types, limiting their broad applicability. This new research aimed to overcome this limitation by creating a single, robust model capable of generalizing across different cancer types, patient populations, sample preparations, and slide scanners.
The development of this universal model was a massive undertaking, utilizing over 20,000 whole-slide images from more than 4,000 patients diagnosed with colorectal, endometrial, lung, or prostate carcinoma. These images were of haematoxylin and eosin (H&E) stained tissue sections from formalin-fixed, paraffin-embedded (FFPE) tissue blocks, a standard in pathology. To ensure its real-world applicability, the model underwent rigorous validation in pre-planned analyses on external cohorts, including over 3,000 additional patients across six cancer types, notably breast and bladder carcinoma, which were not part of the initial training data.
The results were highly encouraging. The model achieved an average Dice coefficient (a measure of segmentation accuracy) of over 80% in all validation cohorts with en bloc resection specimens and in additional exploratory analyses using data from The Cancer Genome Atlas (TCGA). Importantly, the study found no loss of performance when comparing this universal model to models specifically trained on single cancer types. This suggests that the extensive and varied training data allowed the pan-cancer model to learn more general characteristics of cancerous tissues, making it highly adaptable.
The model also demonstrated remarkable robustness. Its performance remained consistent across images acquired from different slide scanners, such as the Aperio AT2 and NanoZoomer XR, and even when evaluated on slides prepared at different external laboratories, as seen with the TCGA cohorts. This adaptability to technical variations is vital for widespread clinical implementation.
However, the study did highlight certain challenges. The model struggled with small, fragmented samples, particularly those from early-stage urothelial carcinoma of the bladder (VUr1 cohort). In these cases, while a positive signal was often present in the probability maps, it was sometimes too weak to be included in the final segmentation. This indicates that detecting very small or fragmented tumor areas remains a hurdle for automated systems.
To provide context, the researchers also compared their model’s performance with MedSAM, a foundation model for medical image segmentation. Their universal model significantly outperformed MedSAM, especially when MedSAM was prompted only by the bounding box of the whole tissue foreground. Even when MedSAM was given the bounding box of the manually segmented tumor, the universal model generally showed superior performance.
Also Read:
- Multimodal AI Model Creates Realistic Pathology Images to Boost Cell Segmentation
- Standardizing Evaluation for Interactive Medical Segmentation Tools
In conclusion, this extensive and rigorous evaluation demonstrates the feasibility of generic tumor segmentation by a single deep learning model across diverse cancer types. This advancement holds immense promise for digital pathology, potentially streamlining diagnostic workflows, assisting pathologists in identifying tumor areas more efficiently, and even extending diagnostic capabilities to rare cancers where specialized models might lack sufficient training data. While challenges with fragmented samples persist, the development of such a generalizable model marks a significant step towards more effective and automated cancer diagnostics. For more details, you can refer to the full research paper here.
