Computer Vision Quality Control for Medical Diagnostics

Overview

Sevamob is an AI-assisted disease management platform that equips frontline health workers with mobile screening tools. Their operational devices analyze clinical imagery across multiple specialties, including sputum microscopy and chest X-rays, to screen for medical conditions.

Bad lighting, blur, and overexposure can easily ruin these images before diagnostic models ever see them. To catch these issues early, we built an offline Quality Control (QC) system that runs directly on mobile hardware. It gives immediate pass/fail feedback before images move to full diagnostic processing.

Challenges of Small Datasets

Modern deep learning models excel on large image datasets, but struggle in clinical quality assurance where data is limited and diagnostic criteria are non-linear.

• Data scarcity: Neural networks need massive datasets to generalize properly. With only a few hundred images available, the models just memorized our training photos instead of learning how to spot defects in unseen ones.
• VLMs: Multimodal models like MedGemma and PaliGemma failed because they were pre-trained for pathology or standard photos, not technical quality checks. Adjusting them required computing power and missed latency targets, while fine-tuning didn’t help much and synthetic data augmentation failed to produce realistic new defect types.
• Prompts: Strict prompts caused models to reject almost every image, whereas softened prompts caused them to miss defective samples entirely (so called confirmation bias).
• Structural defects: While quality assurance often focuses on simple visual artifacts like blur or exposure, our research revealed that the primary cause for sputum slide rejection was presence of saliva in samples, a subtle pattern that standard image filters and neural networks often mistook for normal anatomy.

Lightweight Machine Learning

Instead of forcing heavy AI models into a tight mobile environment, we switched to simple, focused machine learning tools:

• For sputum microscopy, we built a Gradient Boosting classifier operating on 14 fixed mathematical features extracted via OpenCV.
• For chest X-rays, we started from high-dimensional neural networks, but they quickly proved overly complex on the provided dataset. We built a One-Class Isolation Forest algorithm instead, focusing on 5 core optical metrics to catch technical flaws without getting distracted by anatomical noise.
• Additionally to a pass or fail output, we built probability score controls into the inference engine so teams can adjust decision thresholds to their operational environments.

Results

The quality control engine delivered a fast, reliable testing baseline built specifically for mobile hardware:

• Speed: Sputum microscopy response took just 120 ms, which general-purpose AI models cannot achieve on offline mobile hardware.
• Accuracy: The Sputum model hit 83% accuracy on field data, while the X-ray model needs more real-world images to improve its performance.
• 4-Week Delivery: Verified the dataset, selected models, engineered the pipeline, tested and packaged the SDK within the R&D sprint.

Testimonial

“26bitz selected a model, performed training and testing, developed an Android app and did handover with appropriate documentation. They met all project milestones on time. They were very responsive and addressed all our needs.” — Shelley Saxena, Founder & CEO of Sevamob

What’s Next

Sevamob will fine-tune the algorithms as new field samples come in, then integrate the ML engine directly into their mobile devices to test how it handles live clinical workflows.