ObjectiveTo construct an intelligent ultrasound diagnosis system for breast nodules in patients with thyroid dysfunction using deep learning algorithms. MethodsA retrospective analysis was collected breast ultrasound images of 178 patients with thyroid dysfunction from the ultrasound database of the First Affiliated Hospital of Xinjiang Medical University from January 2023 to February 2024, which served as the training set. The deep learning algorithm was used to construct an intelligent ultrasound diagnosis system for breast nodules in patients with thyroid dysfunction. In addition, a retrospective analysis was collected breast ultrasound images of 81 patients with thyroid dysfunction from the ultrasound database of the First Affiliated Hospital of Xinjiang Medical University from March 2024 to January 2025, which served as the validation set. The above system was used as validation set to diagnose whether patients with thyroid dysfunction had breast nodules, and the diagnostic efficacy of imaging physicians’ diagnosis and the intelligent ultrasound diagnosis system for breast nodules in patients with thyroid dysfunction was analyzed. The consistency between the diagnosis of ultrasound physicians, intelligent ultrasound diagnosis system and the “gold standard” was tested by Kappa test. ResultsThere was no statistically significant difference in age, type of thyroid dysfunction, disease duration, number of breast nodules, and other clinical data between the training set and the validation set (P>0.05). The time required for the training set intelligent ultrasound diagnostic system to diagnose a single breast ultrasound image was (0.04±0.01) min, which was lower than that of an ultrasound specialist [(12.36±2.58) min], t=63.709, P＜0.001. The sensitivity, specificity, accuracy, and area under the curve (AUC) of detecting breast nodules in patients with thyroid dysfunction using an intelligent ultrasound diagnostic system were 97.87% (46/47), 100% (34/34), 98.77% (80/81), and 0.997 [95%CI: (0.951, 1.00)], respectively. The sensitivity, specificity, accuracy, and AUC of detecting breast nodules by ultrasound physicians were 89.36% (42/47), 91.18% (31/34), 90.12% (73/81), and 0.904 [95%CI: (0.818, 0.958)], respectively. The AUC of the intelligent ultrasound diagnosis system was higher than that of the ultrasound physician (Z=2.673, P=0.008). The detection results of breast nodules in patients with thyroid dysfunction diagnosed by ultrasound physicians were generally consistent with the “gold standard” (Kappa value=0.799, P＜0.001), while the intelligent ultrasound diagnosis system was in good agreement with the “gold standard” (Kappa value=0.975, P＜0.001). The confusion matrix results showed that the number of false positives was 3 and 0 for the ultrasound department physicians and the intelligent ultrasound diagnostic system, respectively, while the number of false negatives was 5 and 1. The calibration curve results indicated a high consistency between the diagnostic probability and the actual probability of the intelligent ultrasound diagnostic system, with the calibration curve fitting well with the ideal curve (Hosmer-Lemeshow test: χ2=1.246, P=0.997). ConclusionThe intelligent ultrasound diagnosis system for breast nodules in patients with thyroid dysfunction constructed by deep learning algorithm has good diagnostic efficacy, which can help ultrasound physicians improve screening efficiency and accuracy.
Recent years, convolutional neural network (CNN) is a research hot spot in machine learning and has some application value in computer aided diagnosis. Firstly, this paper briefly introduces the basic principle of CNN. Secondly, it summarizes the improvement on network structure from two dimensions of model and structure optimization. In model structure, it summarizes eleven classical models about CNN in the past 60 years, and introduces its development process according to timeline. In structure optimization, the research progress is summarized from five aspects (input layer, convolution layer, down-sampling layer, full-connected layer and the whole network) of CNN. Thirdly, the learning algorithm is summarized from the optimization algorithm and fusion algorithm. In optimization algorithm, it combs the progress of the algorithm according to optimization purpose. In algorithm fusion, the improvement is summarized from five angles: input layer, convolution layer, down-sampling layer, full-connected layer and output layer. Finally, CNN is mapped into the medical image domain, and it is combined with computer aided diagnosis to explore its application in medical images. It is a good summary for CNN and has positive significance for the development of CNN.
It is a significant challenge to improve the blood-brain barrier (BBB) permeability of central nervous system (CNS) drugs in their development. Compared with traditional pharmacokinetic property tests, machine learning techniques have been proven to effectively and cost-effectively predict the BBB permeability of CNS drugs. In this study, we introduce a high-performance BBB permeability prediction model named balanced-stacking-learning based BBB permeability predictor(BSL-B3PP). Firstly, we screen out the feature set that has a strong influence on BBB permeability from the perspective of medicinal chemistry background and machine learning respectively, and summarize the BBB positive(BBB+) quantification intervals. Then, a combination of resampling algorithms and stacking learning(SL) algorithm is used for predicting the BBB permeability of CNS drugs. The BSL-B3PP model is constructed based on a large-scale BBB database (B3DB). Experimental validation shows an area under curve (AUC) of 97.8% and a Matthews correlation coefficient (MCC) of 85.5%. This model demonstrates promising BBB permeability prediction capability, particularly for drugs that cannot penetrate the BBB, which helps reduce CNS drug development costs and accelerate the CNS drug development process.