The development of next-generation sequencing (NGS) has enabled the discovery of cancer-specific driver gene alternations, making precision medicine possible. However, accurate genetic testing requires a sufficient amount of tumor cells in the specimen. The evaluation of tumor content ratio (TCR) from hematoxylin and eosin (H&E)-stained images has been found to vary between pathologists, making it an important challenge to obtain an accurate TCR. In this study, three pathologists exhaustively labeled all cells in 41 regions from 41 lung cancer cases as either tumor, non-tumor or indistinguishable, thus establishing a “gold standard” TCR. We then compared the accuracy of the TCR estimated by 13 pathologists based on visual assessment and the TCR calculated by an AI model that we have developed. It is a compact and fast model that follows a fully convolutional neural network architecture and produces cell detection maps which can be efficiently post-processed to obtain tumor and non-tumor cell counts from which TCR is calculated. Its raw cell detection accuracy is 92% while its classification accuracy is 84%. The results show that the error between the gold standard TCR and the AI calculation was significantly smaller than that between the gold standard TCR and the pathologist’s visual assessment ($p<0.05$). Additionally, the robustness of AI models across institutions is a key issue and we demonstrate that the variation in AI was smaller than that in the average of pathologists when evaluated by institution. These findings suggest that the accuracy of tumor cellularity assessments in clinical workflows is significantly improved by the introduction of robust AI models, leading to more efficient genetic testing and ultimately to better patient outcomes.

Current Vision-Language Models (VLMs) fail at quantitative spatial reasoning because their architectures destroy pixel-level information required for counting and measurements. Vision encoders compress images through patch embeddings, reducing spatial indexing and losing the precise pixel-level tracking required for accurate counting. We present two contributions to address this fundamental limitation. First, we introduce SQuID (Satellite Quantitative Intelligence Dataset), a benchmark of 2,000 satellite image Question-Answer pairs with both numerical range and categorical answers, designed to evaluate quantitative spatial reasoning. The dataset spans three difficulty tiers with annotations automatically generated from human labels and their learned variability. Second, we propose QVLM (Quantitative Vision-Language Model), a code-generation architecture that maintains pixel precision by decoupling language understanding from visual analysis. Instead of encoding images into embeddings, QVLM generates executable code that first calls a segmentation model to obtain pixel-level masks, then operates directly on these masks, preserving spatial indexing throughout the reasoning process. Our experiments show that QVLM using GPT-5 as coder achieves 42.0% accuracy on SQuID compared to 28.1% for a VLM prompted with image-question pairs. Our work reveals that, for quantitative spatial reasoning, architectural decoupling enables better accuracy on quantitative tasks.