Medical image classification using convolutional neural networks (CNNs) is promising but often requires extensive manual tuning for optimal model definition. Neural architecture search (NAS) automates this process, reducing human intervention significantly. This study applies NAS to [18F]-Florbetaben PET cardiac images for classifying cardiac amyloidosis (CA) sub-types (amyloid light chain (AL) and transthyretin amyloid (ATTR)) and controls. Following data preprocessing and augmentation, an evolutionary cell-based NAS approach with a fixed network macro-structure is employed, automatically deriving cells' micro-structure. The algorithm is executed five times, evaluating 100 mutating architectures per run on an augmented dataset of 4048 images (originally 597), totaling 5000 architectures evaluated. The best network (NAS-Net) achieves 76.95% overall accuracy. K-fold analysis yields mean ± SD percentages of sensitivity, specificity, and accuracy on the test dataset: AL subjects (98.7 ± 2.9, 99.3 ± 1.1, 99.7 ± 0.7), ATTR-CA subjects (93.3 ± 7.8, 78.0 ± 2.9, 70.9 ± 3.7), and controls (35.8 ± 14.6, 77.1 ± 2.0, 96.7 ± 4.4). NAS-derived network performance rivals manually determined networks in the literature while using fewer parameters, validating its automatic approach's efficacy.

Brain tumor is a type of disease caused by uncontrolled cell proliferation in the brain leading to serious health issues such as memory loss and motor impairment. Therefore, early diagnosis of brain tumors plays a crucial role to extend the survival of patients. However, given the busy nature of the work of radiologists and aiming to reduce the likelihood of false diagnoses, advancing technologies including computer-aided diagnosis and artificial intelligence have shown an important role in assisting radiologists. In recent years, a number of deep learning-based methods have been applied for brain tumor detection and classification using MRI images and achieved promising results. The main objective of this paper is to present a detailed review of the previous researches in this field. In addition, This work summarizes the existing limitations and significant highlights. The study systematically reviews 60 articles researches published between 2020 and January 2024, extensively covering methods such as transfer learning, autoencoders, transformers, and attention mechanisms. The key findings formulated in this paper provide an analytic comparison and future directions. The review aims to provide a comprehensive understanding of automatic techniques that may be useful for professionals and academic communities working on brain tumor classification and detection.

The differentiation of benign and malignant parotid gland tumors is of major significance as it directly affects the treatment process. In addition, it is also a vital task in terms of early and accurate diagnosis of parotid gland tumors and the determination of treatment planning accordingly. As in other diseases, the differentiation of tumor types involves several challenging, time-consuming, and laborious processes. In the study, Magnetic Resonance (MR) images of 114 patients with parotid gland tumors are used for training and testing purposes by Image Fusion (IF). After the Apparent Diffusion Coefficient (ADC), Contrast-enhanced T1-w (T1C-w), and T2-w sequences are cropped, IF (ADC, T1C-w), IF (ADC, T2-w), IF (T1C-w, T2-w), and IF (ADC, T1C-w, T2-w) datasets are obtained for different combinations of these sequences using a two-dimensional Discrete Wavelet Transform (DWT)-based fusion technique. For each of these four datasets, ResNet18, GoogLeNet, and DenseNet-201 architectures are trained separately, and thus, 12 models are obtained in the study. A Graphical User Interface (GUI) application that contains the most successful of these trained architectures for each data is also designed to support the users. The designed GUI application not only allows the fusing of different sequence images but also predicts whether the label of the fused image is benign or malignant. The results show that the DenseNet-201 models for IF (ADC, T1C-w), IF (ADC, T2-w), and IF (ADC, T1C-w, T2-w) are better than the others, with accuracies of 95.45%, 95.96%, and 92.93%, respectively. It is also noted in the study that the most successful model for IF (T1C-w, T2-w) is ResNet18, and its accuracy is equal to 94.95%.

Natural language processing (NLP) is crucial to extract information accurately from unstructured text to provide insights for clinical decision-making, quality improvement, and medical research. This study compared the performance of a rule-based NLP system and a medical-domain transformer-based model to detect negated concepts in radiology reports. Using a corpus of 984 de-identified radiology reports from a large U.S.-based academic health system (1000 consecutive reports, excluding 16 duplicates), the investigators compared the rule-based medspaCy system and the Clinical Assertion and Negation Classification Bidirectional Encoder Representations from Transformers (CAN-BERT) system to detect negated expressions of terms from RadLex, the Unified Medical Language System Metathesaurus, and the Radiology Gamuts Ontology. Power analysis determined a sample size of 382 terms to achieve α = 0.05 and β = 0.8 for McNemar's test; based on an estimate of 15% negated terms, 2800 randomly selected terms were annotated manually as negated or not negated. Precision, recall, and F1 of the two models were compared using McNemar's test. Of the 2800 terms, 387 (13.8%) were negated. For negation detection, medspaCy attained a recall of 0.795, precision of 0.356, and F1 of 0.492. CAN-BERT achieved a recall of 0.785, precision of 0.768, and F1 of 0.777. Although recall was not significantly different, CAN-BERT had significantly better precision (χ2 = 304.64; p < 0.001). The transformer-based CAN-BERT model detected negated terms in radiology reports with high precision and recall; its precision significantly exceeded that of the rule-based medspaCy system. Use of this system will improve data extraction from textual reports to support information retrieval, AI model training, and discovery of causal relationships.

Our primary aim with this study was to build a patient-level classifier for stroke territory in DWI using AI to facilitate fast triage of stroke to a dedicated stroke center. A retrospective collection of DWI images of 271 and 122 consecutive acute ischemic stroke patients from two centers was carried out. Pretrained MobileNetV2 and EfficientNetB0 architectures were used to classify territorial subtypes as middle cerebral artery, posterior circulation, or watershed infarcts along with normal slices. Various input combinations using edge maps, thresholding, and hard attention versions were explored. The effect of augmenting the three-channel inputs of pre-trained models on classification performance was analyzed. ROC analyses and confusion matrix-derived performance metrics of the models were reported. Of the 271 patients included in this study, 151 (55.7%) were male and 120 (44.3%) were female. One hundred twenty-nine patients had MCA (47.6%), 65 patients had posterior circulation (24%), and 77 patients had watershed (28.0%) infarcts for center 1. Of the 122 patients from center 2, 78 (64%) were male and 44 (34%) were female. Fifty-two patients (43%) had MCA, 51 patients had posterior circulation (42%), and 19 (15%) patients had watershed infarcts. The Mobile-Crop model had the best performance with 0.95 accuracy and a 0.91 mean f1 score for slice-wise classification and 0.88 accuracy on external test sets, along with a 0.92 mean AUC. In conclusion, modified pre-trained models may be augmented with the transformation of images to provide a more accurate classification of affected territory by stroke in DWI.

With the increasing prevalence of mental illness, accurate clinical diagnosis of mental illness is crucial. Compared with MRI, CT has the advantages of wide application, low price, short scanning time, and high patient cooperation. This study aims to construct a deep learning (DL) model based on CT images to make identification of bipolar disorder (BD) and schizophrenia (SZ). A total of 506 patients (BD = 227, SZ = 279) and 179 healthy controls (HC) was collected from January 2022 to May 2023 at two hospitals, and divided into an internal training set and an internal validation set according to a ratio of 4:1. An additional 65 patients (BD = 35, SZ = 30) and 40 HC were recruited from different hospitals, and served as an external test set. All subjects accepted the conventional brain CT examination. The DenseMD model for identify BD and SZ using multiple instance learning was developed and compared with other classical DL models. The results showed that DenseMD performed excellently with an accuracy of 0.745 in the internal validation set, whereas the accuracy of the ResNet-18, ResNeXt-50, and DenseNet-121model was 0.672, 0.664, and 0.679, respectively. For the external test set, DenseMD again outperformed other models with an accuracy of 0.724; however, the accuracy of the ResNet-18, ResNeXt-50, and DenseNet-121model was 0.657, 0.638, and 0.676, respectively. Therefore, the potential of DL models for identification of BD and SZ based on brain CT images was established, and identification ability of the DenseMD model was better than other classical DL models.

This study developed and validated a deep learning-based diagnostic model with uncertainty estimation to aid radiologists in the preoperative differentiation of pathological subtypes of renal cell carcinoma (RCC) based on computed tomography (CT) images. Data from 668 consecutive patients with pathologically confirmed RCC were retrospectively collected from Center 1, and the model was trained using fivefold cross-validation to classify RCC subtypes into clear cell RCC (ccRCC), papillary RCC (pRCC), and chromophobe RCC (chRCC). An external validation with 78 patients from Center 2 was conducted to evaluate the performance of the model. In the fivefold cross-validation, the area under the receiver operating characteristic curve (AUC) for the classification of ccRCC, pRCC, and chRCC was 0.868 (95% CI, 0.826-0.923), 0.846 (95% CI, 0.812-0.886), and 0.839 (95% CI, 0.802-0.88), respectively. In the external validation set, the AUCs were 0.856 (95% CI, 0.838-0.882), 0.787 (95% CI, 0.757-0.818), and 0.793 (95% CI, 0.758-0.831) for ccRCC, pRCC, and chRCC, respectively. The model demonstrated robust performance in predicting the pathological subtypes of RCC, while the incorporated uncertainty emphasized the importance of understanding model confidence. The proposed approach, integrated with uncertainty estimation, offers clinicians a dual advantage: accurate RCC subtype predictions complemented by diagnostic confidence metrics, thereby promoting informed decision-making for patients with RCC.

To establish a reliable machine learning model to predict malignancy in breast lesions identified by ultrasound (US) and optimize the negative predictive value to minimize unnecessary biopsies. We included clinical and ultrasonographic attributes from 1526 breast lesions classified as BI-RADS 3, 4a, 4b, 4c, 5, and 6 that underwent US-guided breast biopsy in four institutions. We selected the most informative attributes to train nine machine learning models, ensemble models and models with tuned threshold to make inferences about the diagnosis of BI-RADS 4a and 4b lesions (validation dataset). We tested the performance of the final model with 403 new suspicious lesions. The most informative attributes were shape, margin, orientation and size of the lesions, the resistance index of the internal vessel, the age of the patient and the presence of a palpable lump. The highest mean negative predictive value (NPV) was achieved with the K-Nearest Neighbors algorithm (97.9%). Making ensembles did not improve the performance. Tuning the threshold did improve the performance of the models and we chose the algorithm XGBoost with the tuned threshold as the final one. The tested performance of the final model was: NPV 98.1%, false negative 1.9%, positive predictive value 77.1%, false positive 22.9%. Applying this final model, we would have missed 2 of the 231 malignant lesions of the test dataset (0.8%). Machine learning can help physicians predict malignancy in suspicious breast lesions identified by the US. Our final model would be able to avoid 60.4% of the biopsies in benign lesions missing less than 1% of the cancer cases.

Malignant cerebral edema (MCE) is a significant complication following endovascular thrombectomy (EVT) in the treatment of acute ischemic stroke. This study aimed to develop and validate a deep learning-assisted diagnosis model based on the hyperattenuated imaging marker (HIM), characterized by hyperattenuation on head non-contrast computed tomography immediately after thrombectomy, to facilitate radiologists in predicting MCE in patients receiving EVT. This study included 271 patients, with 168 in the training cohort, 43 in the validation cohort, and 60 in the prospective internal test cohort. Deep learning models including ResNet 50, ResNet 101, ResNeXt50_32×4d, ResNeXt101_32×8d, and DenseNet 121 were constructed. The performance of senior and junior radiologists with and without optimal model assistance was compared. ResNeXt101_32×8d had the best predictive performance, the analysis of the receiver operating characteristic curve indicated an area under the curve (AUC) of 0.897 for the prediction of MCE in the validation group and an AUC of 0.889 in the test group. Moreover, with the assistance of the model, radiologists exhibited a significant improvement in diagnostic performance, the AUC increased by 0.137 for the junior radiologist and 0.096 for the junior radiologist respectively. Our study utilized the ResNeXt-101 neural network, combined with HIM, to validate a deep learning model for predicting MCE post-EVT. The developed deep learning model demonstrated high discriminative ability, and can serve as a valuable adjunct to radiologists in clinical practice.

Accurately distinguishing histological subtypes and risk categorization of thymomas is difficult. To differentiate the histologic risk categories of thymomas, we developed a combined radiomics model based on non-enhanced and contrast-enhanced computed tomography (CT) radiomics, clinical, and semantic features. In total, 360 patients with pathologically-confirmed thymomas who underwent CT examinations were retrospectively recruited from three centers. Patients were classified using improved pathological classification criteria as low-risk (LRT: types A and AB) or high-risk (HRT: types B1, B2, and B3). The training and external validation sets comprised 274 (from centers 1 and 2) and 86 (center 3) patients, respectively. A clinical-semantic model was built using clinical and semantic variables. Radiomics features were filtered using intraclass correlation coefficients, correlation analysis, and univariate logistic regression. An optimal radiomics model (Rad_score) was constructed using the AutoML algorithm, while a combined model was constructed by integrating Rad_score with clinical and semantic features. The predictive and clinical performances of the models were evaluated using receiver operating characteristic/calibration curve analyses and decision-curve analysis, respectively. Radiomics and combined models (area under curve: training set, 0.867 and 0.884; external validation set, 0.792 and 0.766, respectively) exhibited performance superior to the clinical-semantic model. The combined model had higher accuracy than the radiomics model (0.79 vs. 0.78, p<0.001) in the entire cohort. The original_firstorder_median of venous phase had the highest relative importance among features in the radiomics model. Radiomics and combined radiomics models may serve as noninvasive discrimination tools to differentiate thymoma risk classifications.