Professor UCLA Los Angeles, California, United States
Introduction: : Digital pathology has transformed cancer diagnostics by enabling whole-slide imaging (WSI), remote evaluation, and AI-assisted analysis. However, access to these tools remains relatively limited due to the high cost, large size, and operational demands of commercial scanners, which can cost over $200,000 and weigh upwards of 50 kg (Figure 1A). These limitations make digital pathology inaccessible in many low-resource or decentralized settings, where faster triage and diagnosis could have a significant clinical impact. To address this need, we developed BlurryScope, a compact and affordable scanning microscope that leverages motion-blurred video acquisition and deep learning for rapid tissue classification. Instead of relying on high-end optics or stop-and-stare scanning, BlurryScope continuously acquires blurred images while moving at 5,000 µm/s, reducing mechanical complexity and cost. These compromised images are then processed by a neural network trained to classify HER2 scores—a key biomarker in breast cancer—despite motion blur artifacts. BlurryScope automates the full workflow (Figure 1B) from scanning to HER2 classification using a Fourier transform-based deep learning pipeline. With a component cost under $650 and a total weight of just 2.26 kg, the system offers an accessible alternative or supplement to traditional WSI platforms. This work demonstrates a viable path toward democratizing digital pathology through AI-driven hardware simplification, especially in underserved healthcare environments.
Materials and
Methods: : BlurryScope is a compact scanning microscope constructed using a modified AmScope M150 optical system, a 10×/0.25 NA objective lens, an RGB CMOS camera, and a custom 3D-printed mechanical stage driven by stepper motors. The system operates in a continuous linear scanning mode at 5,000 µm/s, introducing bidirectional motion blur into each stitched image. A full whole-slide scan takes under five minutes, with a throughput of 2.4 mm²/s. The raw video output is processed using automated MATLAB-based stitching algorithms to generate whole-slide images. To demonstrate clinical relevance, we applied BlurryScope to the classification of HER2 scores on immunohistochemically (IHC) stained breast tissue microarrays (TMAs). The dataset included 1,144 patient tissue cores for training and 284 cores for blind testing. Each slide was scanned three times to assess consistency. A Fourier transform-based deep neural network was trained to perform both 4-class (0, 1+, 2+, 3+) and 2-class (0/1+ vs. 2+/3+) HER2 scoring. The full pipeline—scanning, stitching, cropping, labeling, and classification—is fully automated (Figure 1B). Classification models were implemented in PyTorch and optimized with AdamW. Confidence intervals (CIs) were used to filter uncertain predictions. Testing was performed on a workstation equipped with an NVIDIA RTX 3090 GPU. The total system cost remained under $650.
Results, Conclusions, and Discussions:: BlurryScope achieved high performance in automated HER2 scoring, demonstrating classification accuracies of 79.3% for 4-class (0, 1+, 2+, 3+) and 89.7% for 2-class (0/1+ vs. 2+/3+) inference on a blind test set of 284 patient tissue cores. Reproducibility was evaluated by scanning each specimen three times under varying orientations, yielding a prediction consistency of 86.2% across repeated measurements. Accuracy improved further when filtering low-confidence predictions using confidence interval (CI) thresholds, with a minimal trade-off in coverage. Despite acquiring motion-blurred images, the deep learning pipeline successfully extracted diagnostic features relevant for HER2 scoring, validating the viability of AI-compensated hardware simplification. Compared to commercial WSI scanners, BlurryScope maintained competitive accuracy (~5% lag) while operating at a fraction of the cost, weight, and size. These results support the utility of BlurryScope as a triage tool or rapid pre-screening platform in digital pathology. Its low-cost design ( <$650), compact footprint (35×35×35 cm), and end-to-end automation position it as a scalable solution for clinics, labs, or remote environments with limited resources. While not a replacement for gold-standard scanners, BlurryScope offers a practical and accessible alternative that leverages AI to democratize histopathological analysis. Future work will explore multi-biomarker expansion and 3D imaging applications.
