# Preoperative multiclass classification of thymic mass lesions based on radiomics and machine learning

## TL;DR

This study uses CT scans and machine learning to classify thymic mass lesions before surgery, improving early diagnosis and treatment planning.

## Contribution

A machine learning model is developed for preoperative classification of thymic mass lesions using radiomic features and clinical parameters.

## Key findings

- The model achieved an accuracy of 0.8547 in classifying thymic mass lesions.
- Radiomic features from CT scans and age were used to distinguish between thymic cysts and thymomas.

## Abstract

Apart from rare cases such as lymphomas, germ cell tumors, neuroendocrine neoplasms, and thymic hyperplasia, thymic mass lesions (TMLs) are typically categorized into cysts, and thymomas. However, the classification results cannot be determined in advance and can only be confirmed through postoperative pathology. Therefore, the objective of this study is to rely on clinical parameters and radiomic features extracted from chest computed tomography (CT) scans to facilitate the preoperative classification of TMLs. The model development specifically focused on thymic cysts and thymomas, as these are the most commonly encountered anterior mediastinal tumors in clinical practice.

This retrospective study included 400 participants from 3 hospitals between September 2017 and September 2024 due to TMLs. The participants were classified into 7 groups based on the ultimately confirmed etiology: thymic cysts and thymomas, including types A, AB, B1, B2, B3, and C. All participants underwent contrast-enhanced chest CT scans, with senior radiologists delineating regions of interest to extract radiomic features. Additionally, the participants’ ages were also collected as clinical parameters for analysis. The participants were randomly allocated into a training set and a validation set at a 7:3 ratio. A classifier models were developed using the data from the training set, and their performances were evaluated on the validation set.

The model exhibited good classification performance with accuracies of 0.8547.

The model can assist in early diagnosis and the development of personalized treatment strategies for patients with TMLs.

The online version contains supplementary material available at 10.1186/s40644-025-00839-3.