Diagnosis of Mesothelioma Using Image Segmentation and Class-Based Deep Feature Transformations

Our plain-English summary of Diagnosis of Mesothelioma Using Image Segmentation and Class-Based Deep Feature Transformations, by S. Aydın, M. Ağar, M. Çakmak, et al (Diagnostics, 2025).

The study

• Authors: S. Aydın, M. Ağar, M. Çakmak, M. Toğaçar
• Journal: Diagnostics
• Published: 2025;15(18)
• Read it: PubMed · Publisher · Free full text

Abstract

Background/Objectives: Mesothelioma is a rare and aggressive form of cancer that primarily affects the lining of the lungs, abdomen, or heart. It typically arises from exposure to asbestos and is often diagnosed at advanced stages. Limited datasets and complex tissue structures contribute to delays in diagnosis. This study aims to develop a novel hybrid model to improve the accuracy and timeliness of mesothelioma diagnosis. Methods: The proposed approach integrates automatic image segmentation, transformer-based model training, class-based feature extraction, and image transformation techniques. Initially, CT images were processed using the segment anything model (SAM) for region-focused segmentation. These segmented images were then used to train transformer models (CaiT and PVT) to extract class/type-specific features. Each class-based feature set was transformed into an image using Decoder, GAN, and NeRV techniques. Discriminative score and class centroid analysis were then applied to select the most informative image representation for each input. Finally, classification was performed using a residual-based support vector machine (SVM). Results: The proposed hybrid method achieved a classification accuracy of 99.80% in diagnosing mesothelioma, demonstrating its effectiveness in handling limited data and complex tissue characteristics. Conclusions: The results indicate that the proposed model offers a highly accurate and efficient approach to mesothelioma diagnosis. By leveraging advanced segmentation, feature extraction, and representation techniques, it effectively addresses the major challenges associated with early and precise detection of mesothelioma.

In plain English

The authors built and tested an artificial intelligence method for telling mesothelioma apart from benign pleural disease on chest CT scans. They used 1008 CT images from 172 patients at Fırat University Research Hospital in Türkiye. Of these, 504 images came from 86 patients diagnosed with mesothelioma between 2012 and 2024. The other 504 came from 86 patients investigated for suspected mesothelioma who were ultimately found to have benign pleural disease. Metastatic pleural disease was not included. Asbestos exposure was present in 131 (76.2%) of the 172 patients; fiber types and doses were not described. When trained on segmented (outlined) images, the two main models (CaiT and PVT-v2) reached 93.40% and 94.39% accuracy. Classifying individual CT images, the full method reached 99.67% accuracy with a 70/30 training-testing split (bootstrap 95% confidence interval 0.990 to 1.000) and up to 99.80% with five-fold cross-validation (interval 0.995 to 1.000). The authors concluded that the model showed high success in diagnosing mesothelioma. They noted that the data came from one hospital and are not public, and that the method was not tested on outside data.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Retrospective, single-center study developing and testing an AI (deep learning) image-classification model that compares chest CT images of patients with mesothelioma against those of patients with benign pleural disease
• Who was studied: 172 patients at Fırat University Research Hospital, Türkiye: 86 diagnosed with mesothelioma (radiological and pathological findings, 2012 to 2024) and 86 investigated for suspected mesothelioma but diagnosed with benign pleural disease (pleural plaques, infection-related, inflammatory, trauma- or surgery-related findings). The dataset had 1008 CT images, 504 per group, and these images were what the model classified. 98 patients were male and 74 female; mean age was 56.8 years.
• Exposure: Asbestos exposure was not the subject of the study. It was reported as a patient characteristic: present in 131 (76.2%) of the 172 patients, with no fiber types, jobs or doses given. A history of smoking was present in 97 (56.4%).
• Main results:
  • Trained on the segmented CT images, the CaiT model reached 93.40% overall accuracy and the PVT-v2 model 94.39%. On the original, unsegmented images the figures were 85.48% and 85.81%.
  • The full hybrid method reached 99.67% overall accuracy with the holdout approach (70% of images for training, 30% for testing), for both the CaiT-based and PVT-v2-based versions.
  • In Table 7, holdout accuracy was 0.997 with a bootstrap 95% confidence interval of 0.990 to 1.000, for both the CaiT-based and PVT-v2-based versions.
  • With five-fold cross-validation, the PVT-v2-based version reached 99.80% overall accuracy (Table 7: 0.998, 95% CI 0.995 to 1.000) and the CaiT-based version 99.70%.
  • In tests that removed parts of the method (PVT-v2-based), skipping the image-generation step lowered accuracy to 93.73% (holdout) and 95.05% (cross-validation).
• The authors' conclusion: The authors concluded that their findings confirm the validity of the hybrid model and that it shows high success in diagnosing mesothelioma from CT images.
• Limitations the authors state:
  • The data came from a single center, and the dataset is not publicly accessible because of ethical constraints.
  • There was no external validation on independent datasets, which may limit how far the findings generalize.
  • The model has many components, which may make it hard to maintain and develop over time.
  • The multi-layer design can make the model's outputs hard to interpret.
  • Combining several architectures may require high processing power and memory in real-time use.

Topics

Medical terms in this study, explained in our Lexicon: Mesothelioma.

Related studies

• Current evidence base of FDG-PET/CT imaging in the clinical management of malignant pleural mesothelioma: emerging significance of image segmentation and global disease assessment — S. Basu, B. Saboury, et al. (2011)
• Robust Cell Detection and Segmentation for Image Cytometry Reveal Th17 Cell Heterogeneity — T. Tsujikawa, G. Thibault, et al. (2019)
• Accelerating Whole-Body Diffusion-weighted MRI with Deep Learning-based Denoising Image Filters — K. Zormpas-Petridis, N. Tunariu, et al. (2021)
• Asbestos Detection with Fluorescence Microscopy Images and Deep Learning — C. Cai, T. Nishimura, et al. (2021)
• Construction of Asbestos Slate Deep-Learning Training-Data Model Based on Drone Images — S.C. Baek, K.H. Lee, et al. (2023)
• A multi-class predictor based on a probabilistic model: application to gene expression profiling-based diagnosis of thyroid tumors — N. Yukinawa, S. Oba, et al. (2006)

Browse more research in Diagnostics, or return to medical research.