In this study, we examined the correlation between the Ki-67 expression status in breast cancer and the radiomics features derived from images obtained through photoacoustic imaging. Additionally, we developed a column-line diagram model that integrates clinical variables and imaging histological features to enable personalized prediction of Ki-67 expression in preoperative breast cancer patients. Meanwhile, we conducted a comparative analysis of the two column-line diagram models, one based on ultrasound images and the other on photoacoustic images.
The Ki-67 index serves as a pivotal marker for tumor proliferative activity[19], playing a crucial role in subtype identification and predicting treatment outcomes in breast cancer[5, 20]. Elevated levels of Ki-67 expression signify aggressive growth in breast cancer, associated with a heightened risk of recurrence and poorer prognosis[21, 22]. Studies have indicated the utility of monitoring the Ki-67 index in neoadjuvant breast cancer patients to predict the optimal duration of neoadjuvant therapy[23]. Moreover, it serves as an early predictor of treatment efficacy and a prognostic factor for long-term outcomes[24, 25]. Therefore, detecting changes in Ki-67 index is important to improve and personalize the treatment of breast cancer patients. Nevertheless, the dynamic assessment of the Ki-67 index necessitates multiple tumor measurements. Studies have indicated that the inter-observer agreement in immunohistochemistry (IHC) detection of Ki-67 is notably low, and there is a lack of reproducibility among different laboratories[26]. Radiomics, an advancing and emerging field, facilitates the extraction of quantitative, reproducible, and visually inaccessible tumor features from diagnostic images[17]. Moreover, it demonstrates robust capabilities in effectively discriminating between benign and malignant tumors[27, 28]. In comparison to biopsy, radiomics comprehensively captures the heterogeneity across the entire tumor volume, presenting a distinct advantage in the assessment of tumor heterogeneity[29].
Previous investigations into predicting Ki-67 expression levels in breast cancer tumor tissues have predominantly centered on ultrasound (US) and magnetic resonance imaging (MRI)[30, 31]. The reported AUC values for the validation set in these studies were 0.808 and 0.740, respectively. Notably, photoacoustic imaging exhibits distinct imaging characteristics compared to ultrasound and magnetic resonance imaging. No studies to date have utilized radiomics from photoacoustic imaging as a predictor of Ki-67 status in breast cancer. Consequently, employing radiomics based on the photoacoustic imaging features of the breast emerges as a crucial component for the assessment of Ki-67. In this study, we employed a more versatile and cost-effective dual-modality photoacoustic imaging/ultrasound imaging tool to integrate features within the tumor extracted by radiomics and clinical risk factors, resulting in a nomogram model. Compared to the AUC of 0.808 reported in the training set by Liu et al[30], the AUC values for the training and validation sets in our study were 0.904 and 0.890, respectively, indicating superior diagnostic performance. Furthermore, we conducted feature extraction on photoacoustic images and ultrasound images acquired with a dual-modality photoacoustic/ultrasound imaging system to create a nomogram model. The AUC of the nomogram model based on photoacoustic images was 0.871, whereas the AUC of the nomogram model based on ultrasound images was 0.836. This further underscores that photoacoustic imaging-based radiomics is more effective in predicting the expression level of Ki-67 in breast cancer patients.
Our study is subject to several limitations. Firstly, the patient cohort is derived from a single institution, resulting in a restricted sample size. Therefore, future research should aim at expanding the sample size and conducting a multicenter study for external validation to enhance the generalizability of the model. Secondly, the retrospective nature of the study posed challenges in obtaining precise parameters related to the initial machine settings, introducing the potential influence of the machine type on model performance evaluation. Thirdly, the extraction of photoacoustic histological features involves a time-consuming process of tumor contouring. Alternatively, employing deep learning algorithms executed entirely by the machine itself could lead to more accurate and automatic lesion detection and segmentation, thereby improving overall detection results.