Oral squamous cell carcinoma (OSCC) accounts for more than 90% of oral malignancies. Despite numerous advancements in understanding its biology, the mean five-year survival rate of OSCC is still very poor at about 50%, with even lower rates when the disease is detected at later stages. We investigate the use of clinical photographic images taken by common smartphones for the automated detection of OSCC cases and for the identification of suspicious cases mimicking cancer that require an urgent biopsy. We perform a retrospective study on a cohort of 1470 patients drawn from both hospital records and online academic sources. We examine various deep learning methods for the early detection of OSCC cases as well as for the detection of suspicious cases. Our results demonstrate the efficacy of these methods in both tasks, providing a comprehensive understanding of the patient’s condition. When evaluated on holdout data, the model to predict OSCC achieved an AUC of 0.96 (CI: 0.91, 0.98), with a sensitivity of 0.91 and specificity of 0.81. When the data are stratified based on lesion location, we find that our models can provide enhanced accuracy (AUC 1.00) in differentiating specific groups of patients that have lesions in the lingual mucosa, floor of mouth, or posterior tongue. These results underscore the potential of leveraging clinical photos for the timely and accurate identification of OSCC.

Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) is essential for monitoring breast cancer treatment response, yet deep learning progress is limited by the lack of standardized, multi-center datasets. We present BreastDCEDL, a deep learning-ready dataset comprising pretreatment 3D DCE-MRI scans from 2,070 patients across three cohorts: I-SPY1 (n = 172), I-SPY2 (n = 982), and Duke (n = 916). Raw DICOM files were converted to standardized 3D NIfTI format, preserving signal integrity and spatial resolution. The dataset includes unified tumor annotations—binary masks for I-SPY cohorts and bounding boxes for Duke—and harmonized clinical metadata. Pathologic complete response (pCR) data are available for 1,452 patients (100% of I-SPY1/I-SPY2, 32.5% of Duke). Clinical biomarkers include HR status (64.1% positive) and HER2 status (22.2% positive) for over 99% of patients. Predefined train-validation-test splits (1,532/268/270) maintain balanced distributions. BreastDCEDL fills a critical gap in public imaging resources, supporting development of advanced models for pCR prediction and other breast cancer analyses. The dataset is designed for broad reuse and is accompanied by processing tools to support community-driven AI research in oncology.

Background: Papilledema, an ophthalmic finding associated with increased intracranial pressure, is often induced by dermatological medications, including corticosteroids, isotretinoin, and tetracyclines. Early detection is crucial for preventing irreversible optic nerve damage, but access to ophthalmologic expertise is often limited in rural settings. Artificial intelligence (AI) may enable the automated and accurate detection of papilledema from fundus images, thereby supporting timely diagnosis and management. Objective: The primary objective of this study was to explore the diagnostic capability of ChatGPT-4o, a general large language model with multimodal input, in identifying papilledema from fundus photographs. For context, its performance was compared with a ResNet-based convolutional neural network (CNN) specifically fine-tuned for ophthalmic imaging, as well as with the assessments of two human ophthalmologists. The focus was on applications relevant to dermatological care in resource-limited environments. Methods: A dataset of 1094 fundus images (295 papilledema, 799 normal) was preprocessed and partitioned into a training set and a test set. The ResNet model was fine-tuned using discriminative learning rates and a one-cycle learning rate policy. GPT-4o and two human evaluators (a senior ophthalmologist and an ophthalmology resident) independently assessed the test images. Diagnostic metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and Cohen’s Kappa, were calculated for each evaluator. Results: GPT-4o, when applied to papilledema detection, achieved an overall accuracy of 85.9% with substantial agreement beyond chance (Cohen’s Kappa = 0.72), but lower specificity (78.9%) and positive predictive value (73.7%) compared to benchmark models. For context, the ResNet model, fine-tuned for ophthalmic imaging, reached near-perfect accuracy (99.5%, Kappa = 0.99), while two human ophthalmologists achieved accuracies of 96.0% (Kappa ≈ 0.92). Conclusions: This study explored the capability of GPT-4o, a large language model with multimodal input, for detecting papilledema from fundus photographs. GPT-4o achieved moderate diagnostic accuracy and substantial agreement with the ground truth, but it underperformed compared to both a domain-specific ResNet model and human ophthalmologists. These findings underscore the distinction between generalist large language models and specialized diagnostic AI: while GPT-4o is not optimized for ophthalmic imaging, its accessibility, adaptability, and rapid evolution highlight its potential as a future adjunct in clinical screening, particularly in underserved settings. These findings also underscore the need for validation on external datasets and real-world clinical environments before such tools can be broadly implemented.

BACKGROUND Breast cancer is a leading cause of cancer-related mortality worldwide, making accurate treatment response monitoring critical. Artificial intelligence offers unprecedented potential for analyzing complex medical imaging data, yet progress in breast cancer has been limited by the lack of accessible, standardized public datasets. OBJECTIVE To present BreastDCEDL, a comprehensive deep learning-ready dataset for breast Dynamic Contrast-Enhanced MRI (DCE-MRI), and demonstrate its capacity by developing the first transformer-based model for breast DCE-MRI response prediction. METHODS We curated a dataset comprising pretreatment 3D DCE-MRI scans from 2,070 breast cancer patients drawn from the I-SPY 1, I-SPY 2, and Duke cohorts, all sourced from The Cancer Imaging Archive (TCIA). Raw DICOM imaging data was converted into standardized 3D NIfTI volumes, preserving signal integrity. Unified tumor annotations and harmonized clinical metadata, including pathologic complete response (pCR), hormone receptor (HR), and HER2 status, were provided. We implemented a Vision Transformer (ViT) approach with predefined benchmark splits for reproducible research. RESULTS BreastDCEDL consolidates 2,070 patients with preserved original DICOM intensity values, spatial resolution, and imaging parameters in standardized 3D NIfTI format. Our transformer-based model achieved state-of-the-art pCR prediction performance with AUC 0.94 and accuracy 0.93 for HR+/HER2$^-$ tumors, demonstrating the first successful application of transformer models to breast DCE-MRI data. CONCLUSIONS This comprehensive resource addresses the critical need for large-scale, standardized datasets, enabling the development of advanced models including transformer architectures that require substantial training data. BreastDCEDL provides a framework for reproducible research and accelerates progress in AI-assisted breast cancer care.

Breast cancer is the most diagnosed cancer in women, with HER2 status critically guiding treatment decisions. Noninvasive prediction of HER2 status from dynamic contrast-enhanced MRI (DCE-MRI) could streamline diagnostics and reduce reliance on biopsy. However, preprocessing high-dynamic-range DCE-MRI into standardized 8-bit RGB format for pretrained neural networks is nontrivial, and normalization strategy significantly affects model performance. We benchmarked intensity normalization strategies using a Triple-Head Dual-Attention ResNet that processes RGB-fused temporal sequences from three DCE phases. Trained on a multicenter cohort (n=1,149) from the I-SPY trials and externally validated on BreastDCEDL_AMBL (n=43 lesions), our model outperformed transformer-based architectures, achieving 0.75 accuracy and 0.74 AUC on I-SPY test data. N4 bias field correction slightly degraded performance. Without fine-tuning, external validation yielded 0.66 AUC, demonstrating cross-institutional generalizability. These findings highlight the effectiveness of dual-attention mechanisms in capturing transferable spatiotemporal features for HER2 stratification, advancing reproducible deep learning biomarkers in breast cancer imaging.

Background: HER2 status guides targeted therapy in breast cancer but is currently determined by invasive biopsy. Imaging-based HER2 prediction from dynamic contrast-enhanced MRI (DCE-MRI) could provide a non-invasive adjunct decision-support signal, but published models are typically single-center with heterogeneous preprocessing that limits reproducibility. Methods: We trained a Triple-Head Dual-Attention ResNet (THDA-ResNet) that processes three DCE phases (pre-contrast, early post-contrast, and late post-contrast) on the multicenter BreastDCEDL dataset (n = 1149, I-SPY trials), and we compared it with Vision Transformer (ViT) and Convolutional Vision Transformer (CvT) baselines, all ImageNet-pretrained. We benchmarked 14 preprocessing strategies, with and without N4 bias-field correction. External validation used the independent BreastDCEDL_AMBL cohort (43 lesions). AUC confidence intervals used stratified bootstrap; model comparisons used DeLong’s test. Results: THDA-ResNet achieved the highest AUC, 0.74 (95% CI 0.65–0.83), versus 0.66 for ViT and 0.63 for CvT, with the advantage reaching borderline significance over CvT ($p=0.054$) and not significant over ViT ($p=0.14$). At a threshold of 0.7, it retained discrimination (sensitivity 0.41, specificity 0.86), while transformers collapsed to near-trivial classifiers. External AUC was 0.66 (0.49–0.81). N4 correction did not improve performance. Conclusions: Attention-driven CNNs are a strong default for HER2 prediction from DCE-MRI on medium-sized cohorts, and N4 correction can be omitted, simplifying the pipeline.