Siriraj Hospital

INTRODUCTION:Artificial intelligence (AI) and digital pathology have the potential to augment liver biopsy interpretation in MAFLD in clinical practice and trials assessment. However, attitudes and barriers to its implementation have not been systematically explored. METHODS:A survey focusing on conventional liver histology, digital pathology and its AI applications in MAFLD/MASH was conducted among hepatologists and liver pathologists in the Asia Pacific region. RESULTS:AI-assisted digital pathology is perceived to be a valuable addition to existing histological reporting in MAFLD/MASH. Defined standards for application and validation of AI models are important priorities for their implementation. CONCLUSION:There is consensus among clinical experts in the Asia Pacific that AI-assisted histological assessment is useful in MAFLD/MASH interpretation. However, there remain important challenges to the adoption of these technologies into routine clinical workflows.

Purpose The cardiomegaly can be determined by the cardiothoracic ratio (CTR) which can be measured in a chest x-ray image. It is calculated based on a relationship between a size of heart and a transverse dimension of chest. The cardiomegaly is identified when the ratio is larger than a cut-off threshold. This paper aims to propose a solution to calculate the ratio for classifying the cardiomegaly in chest x-ray images. Design/methodology/approach The proposed method begins with constructing lung and heart segmentation models based on U-Net architecture using the publicly available datasets with the groundtruth of heart and lung masks. The ratio is then calculated using the sizes of segmented lung and heart areas. In addition, Progressive Growing of GANs (PGAN) is adopted here for constructing the new dataset containing chest x-ray images of three classes including male normal, female normal and cardiomegaly classes. This dataset is then used for evaluating the proposed solution. Also, the proposed solution is used to evaluate the quality of chest x-ray images generated from PGAN. Findings In the experiments, the trained models are applied to segment regions of heart and lung in chest x-ray images on the self-collected dataset. The calculated CTR values are compared with the values that are manually measured by human experts. The average error is 3.08%. Then, the models are also applied to segment regions of heart and lung for the CTR calculation, on the dataset computed by PGAN. Then, the cardiomegaly is determined using various attempts of different cut-off threshold values. With the standard cut-off at 0.50, the proposed method achieves 94.61% accuracy, 88.31% sensitivity and 94.20% specificity. Originality/value The proposed solution is demonstrated to be robust across unseen datasets for the segmentation, CTR calculation and cardiomegaly classification, including the dataset generated from PGAN. The cut-off value can be adjusted to be lower than 0.50 for increasing the sensitivity. For example, the sensitivity of 97.04% can be achieved at the cut-off of 0.45. However, the specificity is decreased from 94.20% to 79.78%.

Multimodal clinical models are usually judged on accuracy with every modality present, but deployment removes modalities; an echocardiogram is often unavailable where an ECG is routine. Two questions then matter beyond the size of the accuracy loss: which modality was responsible, and whether the model fails loudly or silently once that modality is dropped. The distinction is per-example and modality-level, and is separate from post-hoc feature attribution (e.g. SHAP). Models are replaced often; the evaluation that answers these questions is reused. We present a model-agnostic modality-failure framework: given N modality embeddings, any mask-aware probe, and labels, it returns a per-example failure taxonomy, a per-modality complementarity matrix that attributes error to modalities, and a loud-vs-silent dropout profile separating monitorable failures from those that pass unflagged far from the decision boundary, using only deployment-observable signals. We release it as a small, unit-tested harness and validate it against planted ground truth. Across seeds it recovers that planted modality dominance and complementary subset, reports per-modality loud-vs-silent rates, and scales to a three-modality complementarity matrix; because the planted structure is known by construction, this validates recovery of per-example attribution rather than clinical performance. We then instantiate the framework on frozen EchoJEPA and HuBERT-ECG embeddings for LVEF and the EF <= 40