mesothelioma machine learning approach. It allows for easier detection of mesothelioma than the current, labor-intensive, visual review. In this context, we investigated a machine learning approach and found that neural networks were better at automatically differentiating malignant from benign thoracic lesions. The major advantage of such a learning-based approach is its high capability to recognize complex and sometimes subtle patterns. We focused on the combination of three input classes, namely, texture, shape, and surface-based features, which proved to provide the best discrimination power. Other studies on CT images with pulmonary nodules have also shown that texture analysis combined with shape is the best feature in determining malignancy \[[@CR15]--[@CR17]\]. Furthermore, the combination of radiomics and semantic features to predict malignancy in lymph nodes has been previously investigated \[[@CR18]\]. It is expected that as deep learning approaches improve, they will be able to deal with larger and more diverse datasets and better discriminate between benign and malignant features. There are some limitations that should be considered. First, there were only a limited number of cases in our cohort, especially for the benign category, which reduces the statistical power. Larger datasets are needed to draw solid conclusions. Second, the benign category was heterogeneous, with four different histological types. Although the combination of shape, texture, and shape, with the inclusion of surface-based information led to a strong model, this would not be applicable to a general population, as we could not identify the histological subtypes of some benign cases. A larger sample size with more different benign cases is therefore needed to investigate further these subtypes. Third, a single trained radiologist has diagnosed the cases. The radiologist's level of expertise varies from case to case. However, this study presents a proof of concept of this new classification method. Different radiologists with different levels of expertise can help confirm the generalizability of the findings of our method. Fourth, deep learning may suffer from overfitting when trained with a small number of images. We used a grid search method to avoid this, but overfitting is still possible when using a small dataset. Fifth, our machine learning method might miss some subtle patterns and have a lower diagnostic performance than human experts. For example, the radiologist missed 1 case of benign mesothelioma. The radiologist was very concerned about whether the feature used in our model was a good diagnostic predictor. After reading the images carefully, the radiologist realized that the lesion should have been classified as a malignant case. Radiologists usually diagnose more with their visual expertise and clinical experience. Although we can train and evaluate a model in a short time and using limited computational resources, it is not possible to mimic radiologists' decision processes in general. However, we anticipate that this method is useful for other observers and it is possible to extend the model to predict the histological types of benign lesions. We can assume that radiologists can also be able to train and predict the histological types of benign lesions with training using cases previously reviewed by radiologists. Conclusions {#Sec9} =========== In this paper, we presented a machine learning model to identify benign lesions, which are currently indistinguishable from malignant pleural lesions at early stages. The major advantage of the proposed machine learning method is its capability to learn high-dimensional and complex patterns. The results of our method were compared with those of the visual assessment and manual measurements of radiologists, showing that the algorithm performs better. Therefore, we expect that this system can be used for early detection and screening of pleural mesothelioma with improved speed, cost, and efficiency, making the detection of pleural mesothelioma possible at the early stage, prior to the development of symptoms. AUC : Area under the curve CT : Computed tomography MESO-PRO : Mesothelioma Prognostic Index MR : Magnetic resonance We would like to thank Miss Yuriko Ota and Mrs. Mariko Sugihara for their support in the radiological database search. Funding {#FPar1} ======= The work described was supported by the grant-in-aid for scientific research of the Japanese Society for the Promotion of Science (No. 26242065) and research funding for the National Institute of Radiological Sciences. Availability of data and materials {#FPar2} ================================== All data generated or analyzed during this study are included in this published article (and its supplementary information files). All authors contributed to the study conception and design, data analysis and interpretation, and critical revision of the manuscript. MA was responsible for study design and data collection. YI participated in the study design, supervised the data analysis, and helped draft the manuscript. MS was responsible for data collection, statistical analysis, and drafting of the manuscript. YS was responsible for data collection, statistical analysis, and drafting of the manuscript. MT was responsible for the study conception and data collection. ST performed the interpretation of statistical analysis and participated in drafting the manuscript. All authors read and approved the final manuscript. Ethics approval and consent to participate {#FPar3} ========================================== Not applicable Consent for publication {#FPar4} ======================= Not applicable Competing interests {#FPar5} =================== The authors declare that they have no competing interests. Publisher's Note {#FPar6} ================ Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.