Intelligent Digital and Molecular Pathology

Our AIDP program’s goals include restoring image quality for cytology/pathology samples, developing a framework for AI models such as A!Prostate for Gleason grading, and utilizing generative AI to reduce dependence on human annotations. We also plan to implement autonomous cycled reinforcement learning to improve AI model optimization and generalization. By validating clinical AI pathological diagnostic models through international deployment and local cross-validation, we ensure high performance and accuracy. Our collaboration with institutions like A*STAR, NUH, SGH, and TTSH allows us to develop a comprehensive digital pathology development and adoption platform, integrating AI with a real-world pathological database. This initiative promises to enhance healthcare efficiency, generate commercially appealing intellectual properties, and stimulate job creation and talent development, positioning Singapore as a leading hub for AI-based diagnostics.

Establishing a Comprehensive Pipeline for Medical-Class Digital Pathology Diagnostic Models

The establishment of a comprehensive pathological AIDP model development framework is essential to address the current gaps in standardized AI development pipelines across various diseases. Our innovative framework integrates general AI model optimization, generalization, and validation processes, supported by advanced tools like A!HistoClouds and A!MagQC. This pipeline, spanning from sample preparation to clinical decision-making, includes standard operating protocols (SOPs) for quality control and optimized imaging solutions. Large-scale DP data annotation and server/cloud-based storage bridge the gap between biomedical research and AI fields. Quality control (QC) and generalization are pivotal in this framework. A!MagQC ensures the consistency and reliability of image quality, transforming traditional subjective methods into objective, automated processes. This tool quantitatively assesses common image quality issues, ensuring high-quality data for AI model training and application. A!HistoClouds, on the other hand, facilitates pathologist-AI interaction (PAI), tele-pathology, and assistive pathological diagnosis, accelerating data curation through faster semi-automatic annotation. The development of medical-class R&D pipelines, workflows, and iterative optimization protocols from local and international data flows enhances the generalizability and robustness of AI models, ensuring their applicability across diverse clinical settings. By integrating these advanced QC and generalization methods, our framework aims to produce reliable, efficient, and scalable AI diagnostic solutions, significantly improving healthcare outcomes.

Revolutionizing Quality Control in AI Digital Pathology: The A!MagQC System

We designed A!MagQC to revolutionize image quality assessment, transforming traditional subjective, manual, and qualitative glass-slide QC into an objective, automated, and quantitative system. By applying advanced image processing techniques, A!MagQC comprehensively evaluates digital pathology whole slide images (WSI) across five critical categories: I) Out of focus, II) Contrast, III) Saturation, IV) Texture uniformity, and V) Artifacts. An image patch is deemed “low quality” if A!MagQC detects two or more issues among these categories, subsequently assigning a quantitative score to determine the overall quality of each WSI. A!MagQC has demonstrated remarkable efficiency, significantly reducing annotation costs and manpower effort while ensuring optimal scanner performance. In our program, A!MagQC will serve as the indispensable “gate-keeper” for our large-scale, high-quality annotated database, underscoring the pivotal role of quality control in AI digital pathology.

A!HistoClouds: Revolutionizing AI-Driven Pathological Annotation

We have developed A!HistoClouds, a user-friendly image annotation platform that integrates AI capabilities and Pathologist-AI interaction (PAI), enabling efficient online annotation and tele-pathology. As a cloud-based and local server-compatible solution, A!HistoClouds allows pathologists to work anytime, anywhere, and facilitates seamless collaboration. A!HistoClouds enhances the concept of PAI by bridging pathologists’ expertise and AI, serving as a key component in the evolution of online AI models. Demonstrated in the workflow, PAI refines AI models based on pathologists’ annotations. The platform supports well-defined annotation schemes tailored to different diseases, adhering to WHO/Association standards. Its AI model-agnostic design ensures compatibility with various AI diagnostic models, providing a holistic integration. Built with a human-centric approach, A!HistoClouds post-processes AI output into consumable visual information, streamlining the diagnostic process. The platform offers a highly interactive viewer, enabling pathologists to navigate large digital pathology images (gigapixel) smoothly. Using a pyramid image approach, similar to Google Maps, pathologists can dynamically view, zoom, and draw annotations on pathological images via the A!HistoClouds WSI viewer, enhancing both usability and efficiency.

Enhancing AI Diagnostic Accuracy Through Adaptive Generalization Techniques

Development and training of our AI-enhanced solution began when scanned images were uploaded to A!HistoClouds, where clinical pathologists annotated regions of interest (ROIs) and assigned labels to them (stroma, normal glands, Gleason Pattern 3-5, for example in prostate diagnosis). After cropping the images into small patches and splitting the data into training and testing sets, we trained a convolutional neural network (CNN) using this training dataset to classify image patches into different label categories. On a per-image patch level, the sensitivity, specificity, positive predictive value, and negative predictive value were 93.0%, 92.7%, 96.3%, and 81.5%, respectively. We then visualized the prediction results on whole slide level by applying the model to the entire image. Comparing the generated heat maps and ground truth (pathologists’ annotations and Gleason scores), the predictions made by the AI model were well-matched with the pathologists’ diagnoses. For every different scanner brand, the digitization techniques and parameters are different. Therefore, the images of a single glass slide scanned by different scanners have varied appearances. To overcome this variation issue, we developed an adaptive solution by applying generalization techniques, named as Image Appearance Migration (IAM) to generate consistent results. Our strategy was to standardize images across these scanners and reduce image inconsistencies. Additionally, we further enhanced the generalizability of the models by enlarging the original training dataset using colour augmentation to simulate the appearance variations of different scanners during training. After applying generalization techniques, our AI models demonstrated increased accuracy and consistency among different scanners, potentially other variation in the upstream steps such as staining. Our experimental results indicate that our model is independent of the scanner hardware used, demonstrating its generalization ability across different scanning devices.