Med Image Anal. 2024 Mar 26;95:103145. doi: 10.1016/j.media.2024.103145. Online ahead of print.

ABSTRACT

In recent years, deep learning (DL) has shown great potential in the field of dermatological image analysis. However, existing datasets in this domain have significant limitations, including a small number of image samples, limited disease conditions, insufficient annotations, and non-standardized image acquisitions. To address these shortcomings, we propose a novel framework called DermSynth3D. DermSynth3D blends skin disease patterns onto 3D textured meshes of human subjects using a differentiable renderer and generates 2D images from various camera viewpoints under chosen lighting conditions in diverse background scenes. Our method adheres to top-down rules that constrain the blending and rendering process to create 2D images with skin conditions that mimic in-the-wild acquisitions, ensuring more meaningful results. The framework generates photo-realistic 2D dermatological images and the corresponding dense annotations for semantic segmentation of the skin, skin conditions, body parts, bounding boxes around lesions, depth maps, and other 3D scene parameters, such as camera position and lighting conditions. DermSynth3D allows for the creation of custom datasets for various dermatology tasks. We demonstrate the effectiveness of data generated using DermSynth3D by training DL models on synthetic data and evaluating them on various dermatology tasks using real 2D dermatological images. We make our code publicly available at https://github.com/sfu-mial/DermSynth3D.

PMID:38615432 | DOI:10.1016/j.media.2024.103145

Eur Radiol Exp. 2024 Apr 15;8(1):47. doi: 10.1186/s41747-024-00446-0.

ABSTRACT

BACKGROUND: To investigate the potential of combining compressed sensing (CS) and artificial intelligence (AI), in particular deep learning (DL), for accelerating three-dimensional (3D) magnetic resonance imaging (MRI) sequences of the knee.

METHODS: Twenty healthy volunteers were examined using a 3-T scanner with a fat-saturated 3D proton density sequence with four different acceleration levels (10, 13, 15, and 17). All sequences were accelerated with CS and reconstructed using the conventional and a new DL-based algorithm (CS-AI). Subjective image quality was evaluated by two blinded readers using seven criteria on a 5-point-Likert-scale (overall impression, artifacts, delineation of the anterior cruciate ligament, posterior cruciate ligament, menisci, cartilage, and bone). Using mixed models, all CS-AI sequences were compared to the clinical standard (sense sequence with an acceleration factor of 2) and CS sequences with the same acceleration factor.

RESULTS: 3D sequences reconstructed with CS-AI achieved significantly better values for subjective image quality compared to sequences reconstructed with CS with the same acceleration factor (p ≤ 0.001). The images reconstructed with CS-AI showed that tenfold acceleration may be feasible without significant loss of quality when compared to the reference sequence (p ≥ 0.999).

CONCLUSIONS: For 3-T 3D-MRI of the knee, a DL-based algorithm allowed for additional acceleration of acquisition times compared to the conventional approach. This study, however, is limited by its small sample size and inclusion of only healthy volunteers, indicating the need for further research with a more diverse and larger sample.

TRIAL REGISTRATION: DRKS00024156.

RELEVANCE STATEMENT: Using a DL-based algorithm, 54% faster image acquisition (178 s versus 384 s) for 3D-sequences may be possible for 3-T MRI of the knee.

KEY POINTS: • Combination of compressed sensing and DL improved image quality and allows for significant acceleration of 3D knee MRI. • DL-based algorithm achieved better subjective image quality than conventional compressed sensing. • For 3D knee MRI at 3 T, 54% faster image acquisition may be possible.

PMID:38616220 | DOI:10.1186/s41747-024-00446-0

Sci Rep. 2024 Apr 14;14(1):8624. doi: 10.1038/s41598-024-59218-w.

ABSTRACT

Intelligent detection of athlete behavior is beneficial for guiding sports instruction. Existing mature target detection algorithms provide significant support for this task. However, large-scale target detection algorithms often encounter more challenges in practical application scenarios. We propose SCB-YOLOv5, to detect standardized movements of gymnasts. First, the movements of aerobics athletes were captured, labeled using the labelImg software, and utilized to establish the athlete normative behavior dataset, which was then enhanced by the dataset augmentation using Mosaic9. Then, we improved the YOLOv5 by (1) incorporating the structures of ShuffleNet V2 and convolutional block attention module to reconstruct the Backbone, effectively reducing the parameter size while maintaining network feature extraction capability; (2) adding a weighted bidirectional feature pyramid network into the multiscale feature fusion, to acquire precise channel and positional information through the global receptive field of feature maps. Finally, SCB-YOLOv5 was lighter by 56.9% than YOLOv5. The detection precision is 93.7%, with a recall of 99% and mAP value of 94.23%. This represents a 3.53% improvement compared to the original algorithm. Extensive experiments have verified that our method. SCB-YOLOv5 can meet the requirements for on-site athlete action detection. Our code and models are available at https://github.com/qingDu1/SCB-YOLOv5 .

PMID:38616199 | DOI:10.1038/s41598-024-59218-w

Surgery. 2024 Apr 13:S0039-6060(24)00142-9. doi: 10.1016/j.surg.2024.03.006. Online ahead of print.

ABSTRACT

BACKGROUND: Currently, surgical site infection surveillance relies on labor-intensive manual chart review. Recently suggested solutions involve machine learning to identify surgical site infections directly from the medical record. Deep learning is a form of machine learning that has historically performed better than traditional methods while being harder to interpret. We propose a deep learning model, a long short-term memory network, for the identification of surgical site infection from the medical record with an attention layer for explainability.

METHODS: We retrieved structured data and clinical notes from the University of Utah Health System's electronic health care record for operative events randomly selected for manual chart review from January 2016 to June 2021. Surgical site infection occurring within 30 days of surgery was determined according to the National Surgical Quality Improvement Program definition. We trained the long short-term memory model along with traditional machine learning models for comparison. We calculated several performance metrics from a holdout test set and performed additional analyses to understand the performance of the long short-term memory, including an explainability analysis.

RESULTS: Surgical site infection was present in 4.7% of the total 9,185 operative events. The area under the receiver operating characteristic curve and sensitivity of the long short-term memory was higher (area under the receiver operating characteristic curve: 0.954, sensitivity: 0.920) compared to the top traditional model (area under the receiver operating characteristic curve: 0.937, sensitivity: 0.736). The top 5 features of the long short-term memory included 2 procedure codes and 3 laboratory values.

CONCLUSION: Surgical site infection surveillance is vital for the reduction of surgical site infection rates. Our explainable long short-term memory achieved a comparable area under the receiver operating characteristic curve and greater sensitivity when compared to traditional machine learning methods. With explainable deep learning, automated surgical site infection surveillance could replace burdensome manual chart review processes.

PMID:38616153 | DOI:10.1016/j.surg.2024.03.006

Eur Geriatr Med. 2024 Apr 14. doi: 10.1007/s41999-024-00970-7. Online ahead of print.

ABSTRACT

PURPOSE: The purposes of the study was to describe the degree of agreement between geriatricians with the answers given by an AI tool (ChatGPT) in response to questions related to different areas in geriatrics, to study the differences between specialists and residents in geriatrics in terms of the degree of agreement with ChatGPT, and to analyse the mean scores obtained by areas of knowledge/domains.

METHODS: An observational study was conducted involving 126 doctors from 41 geriatric medicine departments in Spain. Ten questions about geriatric medicine were posed to ChatGPT, and doctors evaluated the AI's answers using a Likert scale. Sociodemographic variables were included. Questions were categorized into five knowledge domains, and means and standard deviations were calculated for each.

RESULTS: 130 doctors answered the questionnaire. 126 doctors (69.8% women, mean age 41.4 [9.8]) were included in the final analysis. The mean score obtained by ChatGPT was 3.1/5 [0.67]. Specialists rated ChatGPT lower than residents (3.0/5 vs. 3.3/5 points, respectively, P < 0.05). By domains, ChatGPT ​​scored better (M: 3.96; SD: 0.71) in general/theoretical questions rather than in complex decisions/end-of-life situations (M: 2.50; SD: 0.76) and answers related to diagnosis/performing of complementary tests obtained the lowest ones (M: 2.48; SD: 0.77).

CONCLUSION: Scores presented big variability depending on the area of knowledge. Questions related to theoretical aspects of challenges/future in geriatrics obtained better scores. When it comes to complex decision-making, appropriateness of the therapeutic efforts or decisions about diagnostic tests, professionals indicated a poorer performance. AI is likely to be incorporated into some areas of medicine, but it would still present important limitations, mainly in complex medical decision-making.

PMID:38615289 | DOI:10.1007/s41999-024-00970-7

Zhong Nan Da Xue Xue Bao Yi Xue Ban. 2024 Jan 28;49(1):58-67. doi: 10.11817/j.issn.1672-7347.2024.230248.

ABSTRACT

OBJECTIVES: Glioblastoma (GBM) and brain metastases (BMs) are the two most common malignant brain tumors in adults. Magnetic resonance imaging (MRI) is a commonly used method for screening and evaluating the prognosis of brain tumors, but the specificity and sensitivity of conventional MRI sequences in differential diagnosis of GBM and BMs are limited. In recent years, deep neural network has shown great potential in the realization of diagnostic classification and the establishment of clinical decision support system. This study aims to apply the radiomics features extracted by deep learning techniques to explore the feasibility of accurate preoperative classification for newly diagnosed GBM and solitary brain metastases (SBMs), and to further explore the impact of multimodality data fusion on classification tasks.

METHODS: Standard protocol cranial MRI sequence data from 135 newly diagnosed GBM patients and 73 patients with SBMs confirmed by histopathologic or clinical diagnosis were retrospectively analyzed. First, structural T1-weight, T1C-weight, and T2-weight were selected as 3 inputs to the entire model, regions of interest (ROIs) were manually delineated on the registered three modal MR images, and multimodality radiomics features were obtained, dimensions were reduced using a random forest (RF)-based feature selection method, and the importance of each feature was further analyzed. Secondly, we used the method of contrast disentangled to find the shared features and complementary features between different modal features. Finally, the response of each sample to GBM and SBMs was predicted by fusing 2 features from different modalities.

RESULTS: The radiomics features using machine learning and the multi-modal fusion method had a good discriminatory ability for GBM and SBMs. Furthermore, compared with single-modal data, the multimodal fusion models using machine learning algorithms such as support vector machine (SVM), Logistic regression, RF, adaptive boosting (AdaBoost), and gradient boosting decision tree (GBDT) achieved significant improvements, with area under the curve (AUC) values of 0.974, 0.978, 0.943, 0.938, and 0.947, respectively; our comparative disentangled multi-modal MR fusion method performs well, and the results of AUC, accuracy (ACC), sensitivity (SEN) and specificity(SPE) in the test set were 0.985, 0.984, 0.900, and 0.990, respectively. Compared with other multi-modal fusion methods, AUC, ACC, and SEN in this study all achieved the best performance. In the ablation experiment to verify the effects of each module component in this study, AUC, ACC, and SEN increased by 1.6%, 10.9% and 15.0%, respectively after 3 loss functions were used simultaneously.

CONCLUSIONS: A deep learning-based contrast disentangled multi-modal MR radiomics feature fusion technique helps to improve GBM and SBMs classification accuracy.

PMID:38615167 | DOI:10.11817/j.issn.1672-7347.2024.230248

NPJ Digit Med. 2024 Apr 13;7(1):96. doi: 10.1038/s41746-024-01090-z.

ABSTRACT

Atrial fibrillation (AF) often escapes detection, given its frequent paroxysmal and asymptomatic presentation. Deep learning of transthoracic echocardiograms (TTEs), which have structural information, could help identify occult AF. We created a two-stage deep learning algorithm using a video-based convolutional neural network model that (1) distinguished whether TTEs were in sinus rhythm or AF and then (2) predicted which of the TTEs in sinus rhythm were in patients who had experienced AF within 90 days. Our model, trained on 111,319 TTE videos, distinguished TTEs in AF from those in sinus rhythm with high accuracy in a held-out test cohort (AUC 0.96 (0.95-0.96), AUPRC 0.91 (0.90-0.92)). Among TTEs in sinus rhythm, the model predicted the presence of concurrent paroxysmal AF (AUC 0.74 (0.71-0.77), AUPRC 0.19 (0.16-0.23)). Model discrimination remained similar in an external cohort of 10,203 TTEs (AUC of 0.69 (0.67-0.70), AUPRC 0.34 (0.31-0.36)). Performance held across patients who were women (AUC 0.76 (0.72-0.81)), older than 65 years (0.73 (0.69-0.76)), or had a CHA2DS2VASc ≥2 (0.73 (0.79-0.77)). The model performed better than using clinical risk factors (AUC 0.64 (0.62-0.67)), TTE measurements (0.64 (0.62-0.67)), left atrial size (0.63 (0.62-0.64)), or CHA2DS2VASc (0.61 (0.60-0.62)). An ensemble model in a cohort subset combining the TTE model with an electrocardiogram (ECGs) deep learning model performed better than using the ECG model alone (AUC 0.81 vs. 0.79, p = 0.01). Deep learning using TTEs can predict patients with active or occult AF and could be used for opportunistic AF screening that could lead to earlier treatment.

PMID:38615104 | DOI:10.1038/s41746-024-01090-z

Nat Commun. 2024 Apr 13;15(1):3211. doi: 10.1038/s41467-024-47461-8.

ABSTRACT

T cells have the ability to eliminate infected and cancer cells and play an essential role in cancer immunotherapy. T cell activation is elicited by the binding of the T cell receptor (TCR) to epitopes displayed on MHC molecules, and the TCR specificity is determined by the sequence of its α and β chains. Here, we collect and curate a dataset of 17,715 αβTCRs interacting with dozens of class I and class II epitopes. We use this curated data to develop MixTCRpred, an epitope-specific TCR-epitope interaction predictor. MixTCRpred accurately predicts TCRs recognizing several viral and cancer epitopes. MixTCRpred further provides a useful quality control tool for multiplexed single-cell TCR sequencing assays of epitope-specific T cells and pinpoints a substantial fraction of putative contaminants in public databases. Analysis of epitope-specific dual α T cells demonstrates that MixTCRpred can identify α chains mediating epitope recognition. Applying MixTCRpred to TCR repertoires from COVID-19 patients reveals enrichment of clonotypes predicted to bind an immunodominant SARS-CoV-2 epitope. Overall, MixTCRpred provides a robust tool to predict TCRs interacting with specific epitopes and interpret TCR-sequencing data from both bulk and epitope-specific T cells.

PMID:38615042 | DOI:10.1038/s41467-024-47461-8

Oral Surg Oral Med Oral Pathol Oral Radiol. 2024 Feb 20:S2212-4403(24)00067-1. doi: 10.1016/j.oooo.2024.02.011. Online ahead of print.

ABSTRACT

OBJECTIVE: This study endeavored to develop a novel, fully automated deep-learning model to determine the topographic relationship between mandibular third molar (MM3) roots and the inferior alveolar canal (IAC) using panoramic radiographs (PRs).

STUDY DESIGN: A total of 1570 eligible subjects with MM3s who had paired PR and cone beam computed tomography (CBCT) from January 2019 to December 2020 were retrospectively collected and randomly grouped into training (80%), validation (10%), and testing (10%) cohorts. The spatial relationship of MM3/IAC was assessed by CBCT and set as the ground truth. MM3-IACnet, a modified deep learning network based on YOLOv5 (You only look once), was trained to detect MM3/IAC proximity using PR. Its diagnostic performance was further compared with dentists, AlexNet, GoogleNet, VGG-16, ResNet-50, and YOLOv5 in another independent cohort with 100 high-risk MM3 defined as root overlapping with IAC on PR.

RESULTS: The MM3-IACnet performed best in predicting the MM3/IAC proximity, as evidenced by the highest accuracy (0.885), precision (0.899), area under the curve value (0.95), and minimal time-spending compared with other models. Moreover, our MM3-IACnet outperformed other models in MM3/IAC risk prediction in high-risk cases.

CONCLUSION: MM3-IACnet model can assist clinicians in MM3s risk assessment and treatment planning by detecting MM3/IAC topographic relationship using PR.

PMID:38614873 | DOI:10.1016/j.oooo.2024.02.011

Oral Surg Oral Med Oral Pathol Oral Radiol. 2024 Feb 20:S2212-4403(24)00069-5. doi: 10.1016/j.oooo.2024.02.010. Online ahead of print.

ABSTRACT

OBJECTIVES: Age and sex characteristics are evident in cephalometric radiographs (CRs), yet their accurate estimation remains challenging due to the complexity of these images. This study aimed to harness deep learning to automate age and sex estimation from CRs, potentially simplifying their interpretation.

STUDY DESIGN: We compared the performance of 4 deep learning models (SVM, R-net, VGG16-SingleTask, and our proposed VGG16-MultiTask) in estimating age and sex from the testing dataset, utilizing a VGG16-based multitask deep learning model on 4,557 CRs. Gradient-weighted class activation mapping (Grad-CAM) was incorporated to identify sex. Performance was assessed using mean absolute error (MAE), specificity, sensitivity, F1 score, and the area under the curve (AUC) in receiver operating characteristic analysis.

RESULTS: The VGG16-MultiTask model outperformed the others, with the lowest MAE (0.864±1.602) and highest sensitivity (0.85), specificity (0.88), F1 score (0.863), and AUC (0.93), demonstrating superior efficacy and robust performance.

CONCLUSIONS: The VGG multitask model demonstrates significant potential in enhancing age and sex estimation from cephalometric analysis, underscoring the role of AI in improving biomedical interpretations.

PMID:38614872 | DOI:10.1016/j.oooo.2024.02.010

Clin Radiol. 2024 Mar 19:S0009-9260(24)00145-4. doi: 10.1016/j.crad.2024.03.007. Online ahead of print.

ABSTRACT

BACKGROUND: Several studies have been published comparing deep learning (DL)/machine learning (ML) to radiologists in differentiating PCNSLs from GBMs with equivocal results. We aimed to perform this meta-analysis to evaluate the diagnostic accuracy of ML/DL versus radiologists in classifying PCNSL versus GBM using MRI.

METHODOLOGY: The study was performed in accordance with PRISMA guidelines. Data was extracted and interpreted by two researchers with 12 and 23 years' experience, respectively, and QUADAS-2 tool was used for quality and risk-bias assessment. We constructed contingency tables to derive sensitivity, specificity accuracy, summary receiver operating characteristic (SROC) curve, and the area under the curve (AUC).

RESULTS: Our search identified 11 studies, of which 8 satisfied our inclusion criteria and restricted the analysis in each study to reporting the model showing highest accuracy, with a total sample size of 1159 patients. The random effects model showed a pooled sensitivity of 0.89 [95% CI:0.84-0.92] for ML and 0.82 [95% CI:0.76-0.87] for radiologists. Pooled specificity was 0.88 [95% CI: 0.84-0.91] for ML and 0.90 [95% CI: 0.81-0.95] for radiologists. Pooled accuracy was 0.88 [95% CI: 0.86-0.90] for ML and 0.86 [95% CI: 0.78-0.91] for radiologists. Pooled AUC of ML was 0.94 [95% CI:0.92-0.96]and for radiologists, it was 0.90 [95% CI: 0.84-0.93].

CONCLUSIONS: MRI-based ML/DL techniques can complement radiologists to improve the accuracy of classifying GBMs from PCNSL, possibly reduce the need for a biopsy, and avoid any unwanted neurosurgical resection of a PCNSL.

PMID:38614870 | DOI:10.1016/j.crad.2024.03.007

Clin Radiol. 2024 Mar 26:S0009-9260(24)00141-7. doi: 10.1016/j.crad.2024.03.002. Online ahead of print.

ABSTRACT

AIM: To evaluate whether deep learning reconstruction (DLR) can accelerate the acquisition of magnetic resonance imaging (MRI) sequences of the knee for clinical use.

MATERIALS AND METHODS: Using a 1.5-T MRI scanner, sagittal fat-suppressed T2-weighted imaging (fs-T2WI), coronal proton density-weighted imaging (PDWI), and coronal T1-weighted imaging (T1WI) were performed. DLR was applied to images with a number of signal averages (NSA) of 1 to obtain 1DLR images. Then 1NSA, 1DLR, and 4NSA images were compared subjectively, and by noise (standard deviation of intra-articular water or medial meniscus) and contrast-to-noise ratio between two anatomical structures or between an anatomical structure and intra-articular water.

RESULTS: Twenty-seven healthy volunteers (age: 40.6 ± 11.9 years) were enrolled. Three 1DLR image sequences were obtained within 200 s (approximately 12 minutes for 4NSA image). According to objective evaluations, PDWI 1DLR images showed the smallest noise and significantly higher contrast than 1NSA and 4NSA images. For fs-T2WI, smaller noise and higher contrast were observed in the order of 4NSA, 1DLR, and 1NSA images. According to the subjective analysis, structure visibility, image noise, and overall image quality were significantly better for PDWI 1DLR than 1NSA images; moreover, the visibility of the meniscus and bone, image noise, and overall image quality were significantly better for 1DLR than 4NSA images. Fs-T2WI and T1WI 1DLR images showed no difference between 1DLR and 4NSA images.

CONCLUSION: Compared to PDWI 4NSA images, PDWI 1DLR images were of higher quality, while the quality of fs-T2WI and T1WI 1DLR images was similar to that of 4NSA images.

PMID:38614869 | DOI:10.1016/j.crad.2024.03.002

Acad Radiol. 2024 Apr 13:S1076-6332(24)00197-1. doi: 10.1016/j.acra.2024.03.033. Online ahead of print.

ABSTRACT

RATIONALE AND OBJECTIVES: To systematically evaluate the application value of radiomics and deep learning (DL) in the differential diagnosis of benign and malignant soft tissue tumors (STTs).

MATERIALS AND METHODS: A systematic review was conducted on studies published up to December 11, 2023, that utilized radiomics and DL methods for the diagnosis of STTs. The methodological quality and risk of bias were evaluated using the Radiomics Quality Score (RQS) 2.0 system and Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool, respectively. A bivariate random-effects model was used to calculate the summarized sensitivity and specificity. To identify factors contributing to heterogeneity, meta-regression and subgroup analyses were performed to assess the following covariates: diagnostic modality, region/volume of interest, imaging examination, study design, and pathology type. The asymmetry of Deeks' funnel plot was used to assess publication bias.

RESULTS: A total of 21 studies involving 3866 patients were included, with 13 studies using independent test/validation sets included in the quantitative statistical analysis. The average RQS was 21.31, with substantial or near-perfect inter-rater agreement. The combined sensitivity and specificity were 0.84 (95% CI: 0.76-0.89) and 0.88 (95% CI: 0.69-0.96), respectively. Meta-regression and subgroup analyses showed that study design and the region/volume of interest were significant factors affecting study heterogeneity (P < 0.05). No publication bias was observed.

CONCLUSION: Radiomics and DL can accurately distinguish between benign and malignant STTs. Future research should concentrate on enhancing the rigor of study designs, conducting multicenter prospective validations, amplifying the interpretability of DL models, and integrating multimodal data to elevate the diagnostic accuracy and clinical utility of soft tissue tumor assessments.

PMID:38614826 | DOI:10.1016/j.acra.2024.03.033

Eur J Vasc Endovasc Surg. 2024 Apr 11:S1078-5884(24)00340-X. doi: 10.1016/j.ejvs.2024.04.004. Online ahead of print.

NO ABSTRACT

PMID:38614229 | DOI:10.1016/j.ejvs.2024.04.004

Bioinformatics. 2024 Apr 13:btae204. doi: 10.1093/bioinformatics/btae204. Online ahead of print.

ABSTRACT

MOTIVATION: Personalized cancer treatments require accurate drug response predictions. Existing deep learning methods show promise but higher accuracy is needed to serve the purpose of precision medicine. The prediction accuracy can be improved with not only topology but geometrical information of drugs.

RESULTS: A novel deep learning methodology for drug response prediction is presented, named Hi-GeoMVP. It synthesizes hierarchical drug representation with multi-omics data, leveraging graph neural networks and variational autoencoders for detailed drug and cell line representations. Multi-task learning is employed to make better prediction, while both 2D and 3D molecular representations capture comprehensive drug information. Testing on the GDSC dataset confirms Hi-GeoMVP's enhanced performance, surpassing prior state-of-the-art methods by improving the Pearson correlation coefficient from 0.934 to 0.941 and decreasing the root mean square error from 0.969 to 0.931. In the case of blind test, Hi-GeoMVP demonstrated robustness, outperforming the best previous models with a superior Pearson correlation coefficient in the drug-blind test. These results underscore Hi-GeoMVP's capabilities in drug response prediction, implying its potential for precision medicine.

AVAILABILITY AND IMPLEMENTATION: The source code is available at https://github.com/matcyr/Hi-GeoMVP.

SUPPLEMENTARY INFORMATION: Supplementary data is available at Bioinformatics online.

PMID:38614131 | DOI:10.1093/bioinformatics/btae204

Comput Biol Med. 2024 Apr 6;174:108420. doi: 10.1016/j.compbiomed.2024.108420. Online ahead of print.

ABSTRACT

BACKGROUND AND OBJECTIVE: Liver tumor segmentation (LiTS) accuracy on contrast-enhanced computed tomography (CECT) images is higher than that on non-contrast computed tomography (NCCT) images. However, CECT requires contrast medium and repeated scans to obtain multiphase enhanced CT images, which is time-consuming and cost-increasing. Therefore, despite the lower accuracy of LiTS on NCCT images, which still plays an irreplaceable role in some clinical settings, such as guided brachytherapy, ablation, or evaluation of patients with renal function damage. In this study, we intend to generate enhanced high-contrast pseudo-color CT (PCCT) images to improve the accuracy of LiTS and RECIST diameter measurement on NCCT images.

METHODS: To generate high-contrast CT liver tumor region images, an intensity-based tumor conspicuity enhancement (ITCE) model was first developed. In the ITCE model, a pseudo color conversion function from an intensity distribution of the tumor was established, and it was applied in NCCT to generate enhanced PCCT images. Additionally, we design a tumor conspicuity enhancement-based liver tumor segmentation (TCELiTS) model, which was applied to improve the segmentation of liver tumors on NCCT images. The TCELiTS model consists of three components: an image enhancement module based on the ITCE model, a segmentation module based on a deep convolutional neural network, and an attention loss module based on restricted activation. Segmentation performance was analyzed using the Dice similarity coefficient (DSC), sensitivity, specificity, and RECIST diameter error.

RESULTS: To develop the deep learning model, 100 patients with histopathologically confirmed liver tumors (hepatocellular carcinoma, 64 patients; hepatic hemangioma, 36 patients) were randomly divided into a training set (75 patients) and an independent test set (25 patients). Compared with existing tumor automatic segmentation networks trained on CECT images (U-Net, nnU-Net, DeepLab-V3, Modified U-Net), the DSCs achieved on the enhanced PCCT images are both improved compared with those on NCCT images. We observe improvements of 0.696-0.713, 0.715 to 0.776, 0.748 to 0.788, and 0.733 to 0.799 in U-Net, nnU-Net, DeepLab-V3, and Modified U-Net, respectively, in terms of DSC values. In addition, an observer study including 5 doctors was conducted to compare the segmentation performance of enhanced PCCT images with that of NCCT images and showed that enhanced PCCT images are more advantageous for doctors to segment tumor regions. The results showed an accuracy improvement of approximately 3%-6%, but the time required to segment a single CT image was reduced by approximately 50 %.

CONCLUSIONS: Experimental results show that the ITCE model can generate high-contrast enhanced PCCT images, especially in liver regions, and the TCELiTS model can improve LiTS accuracy in NCCT images.

PMID:38613896 | DOI:10.1016/j.compbiomed.2024.108420

Comput Biol Med. 2024 Apr 9;174:108464. doi: 10.1016/j.compbiomed.2024.108464. Online ahead of print.

ABSTRACT

Pulmonary Embolisms (PE) represent a leading cause of cardiovascular death. While medical imaging, through computed tomographic pulmonary angiography (CTPA), represents the gold standard for PE diagnosis, it is still susceptible to misdiagnosis or significant diagnosis delays, which may be fatal for critical cases. Despite the recently demonstrated power of deep learning to bring a significant boost in performance in a wide range of medical imaging tasks, there are still very few published researches on automatic pulmonary embolism detection. Herein we introduce a deep learning based approach, which efficiently combines computer vision and deep neural networks for pulmonary embolism detection in CTPA. Our method brings novel contributions along three orthogonal axes: (1) automatic detection of anatomical structures; (2) anatomical aware pretraining, and (3) a dual-hop deep neural net for PE detection. We obtain state-of-the-art results on the publicly available multicenter large-scale RSNA dataset.

PMID:38613894 | DOI:10.1016/j.compbiomed.2024.108464

Comput Biol Med. 2024 Apr 9;174:108430. doi: 10.1016/j.compbiomed.2024.108430. Online ahead of print.

ABSTRACT

BACKGROUND: To investigate the effectiveness of contrastive learning, in particular SimClr, in reducing the need for large annotated ultrasound (US) image datasets for fetal standard plane identification.

METHODS: We explore SimClr advantage in the cases of both low and high inter-class variability, considering at the same time how classification performance varies according to different amounts of labels used. This evaluation is performed by exploiting contrastive learning through different training strategies. We apply both quantitative and qualitative analyses, using standard metrics (F1-score, sensitivity, and precision), Class Activation Mapping (CAM), and t-Distributed Stochastic Neighbor Embedding (t-SNE).

RESULTS: When dealing with high inter-class variability classification tasks, contrastive learning does not bring a significant advantage; whereas it results to be relevant for low inter-class variability classification, specifically when initialized with ImageNet weights.

CONCLUSIONS: Contrastive learning approaches are typically used when a large number of unlabeled data is available, which is not representative of US datasets. We proved that SimClr either as pre-training with backbone initialized via ImageNet weights or used in an end-to-end dual-task may impact positively the performance over standard transfer learning approaches, under a scenario in which the dataset is small and characterized by low inter-class variability.

PMID:38613892 | DOI:10.1016/j.compbiomed.2024.108430

Comput Biol Med. 2024 Apr 2;174:108387. doi: 10.1016/j.compbiomed.2024.108387. Online ahead of print.

ABSTRACT

Accurate segmentation and lesion localization are essential for treating diseases in medical images. Despite deep learning methods enhancing segmentation, they still have limitations due to convolutional neural networks' inability to capture long-range feature dependencies. The self-attention mechanism in Transformers addresses this drawback, but high-resolution images present computational complexity. To improve the convolution and Transformer, we suggest a hierarchical hybrid multiaxial attention mechanism called H2MaT-Unet. This approach combines hierarchical post-feature data and applies the multiaxial attention mechanism to the feature interactions. This design facilitates efficient local and global interactions. Furthermore, we introduce a Spatial and Channel Reconstruction Convolution (ScConv) module to enhance feature aggregation. The paper introduces the H2MaT-UNet model which achieves 87.74% Dice in the multi-target segmentation task and 87.88% IOU in the single-target segmentation task, surpassing current popular models and accomplish a new SOTA. H2MaT-UNet synthesizes multi-scale feature information during the layering stage and utilizes a multi-axis attention mechanism to amplify global information interactions in an innovative manner. This re-search holds value for the practical application of deep learning in clinical settings. It allows healthcare providers to analyze segmented details of medical images more quickly and accurately.

PMID:38613886 | DOI:10.1016/j.compbiomed.2024.108387

MAGMA. 2024 Apr 13. doi: 10.1007/s10334-024-01156-9. Online ahead of print.

ABSTRACT

PURPOSE: Use a conference challenge format to compare machine learning-based gamma-aminobutyric acid (GABA)-edited magnetic resonance spectroscopy (MRS) reconstruction models using one-quarter of the transients typically acquired during a complete scan.

METHODS: There were three tracks: Track 1: simulated data, Track 2: identical acquisition parameters with in vivo data, and Track 3: different acquisition parameters with in vivo data. The mean squared error, signal-to-noise ratio, linewidth, and a proposed shape score metric were used to quantify model performance. Challenge organizers provided open access to a baseline model, simulated noise-free data, guides for adding synthetic noise, and in vivo data.

RESULTS: Three submissions were compared. A covariance matrix convolutional neural network model was most successful for Track 1. A vision transformer model operating on a spectrogram data representation was most successful for Tracks 2 and 3. Deep learning (DL) reconstructions with 80 transients achieved equivalent or better SNR, linewidth and fit error compared to conventional 320 transient reconstructions. However, some DL models optimized linewidth and SNR without actually improving overall spectral quality, indicating a need for more robust metrics.

CONCLUSION: DL-based reconstruction pipelines have the promise to reduce the number of transients required for GABA-edited MRS.

PMID:38613715 | DOI:10.1007/s10334-024-01156-9