Spondylolisthesis, a condition characterized by vertebral slippage, often results in pain and limited mobility. To enhance the detection of spondylolisthesis in X-ray images, we developed a YOLOv8-based model trained on a dataset of 10,616 images (AP and LA views). To address the variability in X-ray image quality, caused by factors such as machine variations, medical expertise, patient movement, and artifacts, we employed a comprehensive suite of image processing techniques to simulate a wide range of real-world scenarios. These techniques included histogram equalization to adjust image contrast, blurring to simulate motion artifacts, brightness and contrast adjustments to mimic machine settings, negative color transformations to replicate radiologist-specific viewing preferences, and power law transformations to simulate scenarios of improper patient positioning or machine configuration. By augmenting the dataset with these enhanced images, we significantly improved the model’s ability to generalize and accurately detect spondylolisthesis in diverse clinical settings. Our model achieved the highest precision of 96.20% for five classes in AP images without image enhancement and 96.77% for six classes in LA images with enhancement. Furthermore, it exhibited a minimum Euclidean Distance of 0.1992 in five classes AP images without enhancement and 0.1695 in six classes LA images with enhancement. Finally, the model achieved the best Intersection over Union (IOU) of 94.15% in five classes AP images without enhancement and 95.09% in six classes LA images with enhancement.
This research presents an intelligent defect inspection model that uses object detection to identify defects in Flip Chip cross-section images, aiding failure analysis (FA) engineers working with Chip-on-Wafer (CoW) products. The model aims to enhance accuracy, save time and costs, reduce human error, and ensure reliability. The dataset was sourced from an electronics service manufacturer in Thailand and categorized into four classes: good bump, head-in-pillow (HIP) defect, non-wetting defect, and solder void defect. High-resolution images were captured using an Olympus BX53M microscope at 1000× magnification, focusing on a 50-micrometer copper pillar (CP) bump diameter. To address dataset imbalances, we established seven experimental conditions using various augmentation techniques and generative artificial intelligence (AI) to synthesize underrepresented image classes. We evaluated YOLOv5, YOLOv6, YOLOv7, and YOLOv8 across 26 trained models. The training durations for YOLOv5 and YOLOv8 were 0.86 hours, demonstrating a significant time advantage over other versions. F1-score evaluations indicated that YOLOv5 achieved scores between 0.948 and 0.981, outperforming other versions and even inspectors with over 20 years of experience.
Automating ICD-10 coding from discharge summaries remains demanding because coders analyze clinical narratives while justifying decisions. This study compares three automation patterns: PLM-ICD as a standalone deep learning system emitting 15 codes per case, LLM-only generation with full autonomy, and a hybrid approach where PLM-ICD drafts candidates for an agentic LLM filter to accept or reject. All strategies were evaluated on 19,801 MIMIC-IV summaries using four LLMs spanning compact (Qwen2.5-3B, Llama-3.2-3B, Phi-4-mini) through large scale (Sonnet-4.5). Precision guided evaluation because coders still supply any missing diagnoses. PLM-ICD alone reached 55.8% precision while always surfacing 15 suggestions. LLM-only generation lagged severely (1.5--34.6% precision) and produced inconsistent output sizes. The agentic filter delivered the best trade-off: compact LLMs reviewed the 15 candidates, discarded weak evidence, and returned 2--8 high-confidence codes. Llama-3.2-3B, for example, improved from 1.5% as a generator to 55.1% as a verifier while trimming false positives by 73%. These results show that positioning LLMs as quality controllers, rather than primary generators, yields reliable support for clinical coding teams, while formal recall/F1 reporting remains future work for fully autonomous implementations.