LAUSR

Simple Summary Tumor-infiltrating lymphocytes (TILs) have been proven to be promising biomarkers associated with therapeutic outcomes and prognosis in breast cancer patients. Increased TIL levels predicted a higher rate of response to neoadjuvant chemotherapy in all molecular subtypes and was also associated with a survival benefit in human epidermal growth factor receptor 2-positive and triple-negative breast cancer. The assessment of TILs was based on surgical pathological sections or needle biopsies; this process was invasive and may have introduced sampling bias in biopsies. Imaging-based biomarkers provide a non-invasive evaluation of TIL levels. The aim of this study was to explore the feasibility of transformer-based or convolutional neural network (CNN)-based deep-learning (DL) models to predict TIL levels in breast cancer from ultrasound (US) images. We confirmed that the ultrasound-based DL approach was a good non-invasive tool for predicting TILs in breast cancer and provided key complementary information in equivocal cases that were prone to sampling bias. Abstract This study aimed to explore the feasibility of using a deep-learning (DL) approach to predict TIL levels in breast cancer (BC) from ultrasound (US) images. A total of 494 breast cancer patients with pathologically confirmed invasive BC from two hospitals were retrospectively enrolled. Of these, 396 patients from hospital 1 were divided into the training cohort (n = 298) and internal validation (IV) cohort (n = 98). Patients from hospital 2 (n = 98) were in the external validation (EV) cohort. TIL levels were confirmed by pathological results. Five different DL models were trained for predicting TIL levels in BC using US images from the training cohort and validated on the IV and EV cohorts. The overall best-performing DL model, the attention-based DenseNet121, achieved an AUC of 0.873, an accuracy of 79.5%, a sensitivity of 90.7%, a specificity of 65.9%, and an F1 score of 0.830 in the EV cohort. In addition, the stratified analysis showed that the DL models had good discrimination performance of TIL levels in each of the molecular subgroups. The DL models based on US images of BC patients hold promise for non-invasively predicting TIL levels and helping with individualized treatment decision-making.