Prediction of Prostate Cancer Risk Stratification Based on A Nonlinear Transformation Stacking Learning Strategy.

PURPOSE: Prostate cancer (PCa) is an epithelial malignancy that originates in the prostate gland and is generally categorized into low, intermediate, and high-risk groups. The primary diagnostic indicator for PCa is the measurement of serum prostate-specific antigen (PSA) values. However, reliance on PSA levels can result in false positives, leading to unnecessary biopsies and an increased risk of invasive injuries. Therefore, it is imperative to develop an efficient and accurate method for PCa risk stratification. Many recent studies on PCa risk stratification based on clinical data have employed a binary classification, distinguishing between low to intermediate and high risk. In this paper, we propose a novel machine learning (ML) approach utilizing a stacking learning strategy for predicting the tripartite risk stratification of PCa. METHODS: Clinical records, featuring attributes selected using the lasso method, were utilized with 5 ML classifiers. The outputs of these classifiers underwent transformation by various nonlinear transformers and were then concatenated with the lasso-selected features, resulting in a set of new features. A stacking learning strategy, integrating different ML classifiers, was developed based on these new features. RESULTS: Our proposed approach demonstrated superior performance, achieving an accuracy of 0.83 and an area under the receiver operating characteristic curve value of 0.88 in a dataset comprising 197 PCa patients with 42 clinical characteristics. CONCLUSION: This study aimed to improve clinicians' ability to rapidly assess PCa risk stratification while reducing the burden on patients. This was achieved by using artificial intelligence-related technologies as an auxiliary method for diagnosing PCa.

Multimodality deep learning radiomics nomogram for preoperative prediction of malignancy of breast cancer: a multicenter study.

Background. Breast cancer is the most prevalent cancer diagnosed in women worldwide. Accurately and efficiently stratifying the risk is an essential step in achieving precision medicine prior to treatment. This study aimed to construct and validate a nomogram based on radiomics and deep learning for preoperative prediction of the malignancy of breast cancer (MBC).Methods. The clinical and ultrasound imaging data, including brightness mode (B-mode) and color Doppler flow imaging, of 611 breast cancer patients from multiple hospitals in China were retrospectively analyzed. Patients were divided into one primary cohort (PC), one validation cohort (VC) and two test cohorts (TC1 and TC2). A multimodality deep learning radiomics nomogram (DLRN) was constructed for predicting the MBC. The performance of the proposed DLRN was comprehensively assessed and compared with three unimodal models via the calibration curve, the area under the curve (AUC) of receiver operating characteristics and the decision curve analysis.Results. The DLRN discriminated well between the MBC in all cohorts [overall AUC (95% confidence interval): 0.983 (0.973-0.993), 0.972 (0.952-0.993), 0.897 (0.823-0.971), and 0.993 (0.977-1.000) on the PC, VC, test cohorts1 (TC1) and test cohorts2 TC2 respectively]. In addition, the DLRN performed significantly better than three unimodal models and had good clinical utility.Conclusion. The DLRN demonstrates good discriminatory ability in the preoperative prediction of MBC, can better reveal the potential associations between clinical characteristics, ultrasound imaging features and disease pathology, and can facilitate the development of computer-aided diagnosis systems for breast cancer patients. Our code is available publicly in the repository athttps://github.com/wupeiyan/MDLRN.

ULS4US: universal lesion segmentation framework for 2D ultrasound images.

Objective. Deep learning (DL) methods have been widely utilized in ultrasound (US) image segmentation tasks. However, current DL segmentation methods for US images are typically developed only for lesion segmentation of specific organs; e.g. breast or thyroid US. So far, there is currently no general-purpose lesion segmentation framework for US images that can be implemented across various organs in computer aided diagnosis scenarios. Considering that most lesion locations in US images have abnormal ultrasonic echo intensities or patterns that may be visually distinct from surrounding normal tissues or organs, it is thus possible to develop a universal lesion segmentation framework for US images (named as ULS4US), focusing on effectively identifying and segmenting lesions of various sizes in different organs.Approach. The proposed ULS4US framework comprises three components: (1) a multiple-in multi-out (MIMO) UNet that incorporates multiscale features extracted from the US image and lesion, (2) a novel two-stage lesion-aware learning algorithm that recursively locates and segments the lesions in a reinforced manner, and (3) a lesion-adaptive loss function for the MIMO-UNet that integrates two weighted components and one self-supervised component designed for intra- and inter-branches of network outputs, respectively.Main Results. Compared to six state-of-the-art segmentation models, ULS4US has achieved superior performance (accuracy of 0.956, DSC of 0.836, HD of 7.849, and mIoU of 0.731) in a unified dataset consisting of two public and three private US image datasets, which include over 2200 images of three specific types of organs. Comparative experiments on both individual and unified datasets suggest that ULS4US is likely scalable with additional data.Significance. The study demonstrates the potential of DL-based universal lesion segmentation approaches in clinical US, which would substantially reduce clinician workload and enhance diagnostic accuracy.