Eric Cosatto

Peter A. Massih, Eric Cosatto

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.

Eric Cosatto, Kyle Gerard, Hans-Peter Graf et al.

We propose a method to accurately obtain the ratio of tumor cells over an entire histological slide. We use deep fully convolutional neural network models trained to detect and classify cells on images of H&E-stained tissue sections. Pathologists' labels consisting of exhaustive nuclei locations and tumor regions were used to trained the model in a supervised fashion. We show that combining two models, each working at a different magnification allows the system to capture both cell-level details and surrounding context to enable successful detection and classification of cells as either tumor-cell or normal-cell. Indeed, by conditioning the classification of a single cell on a multi-scale context information, our models mimic the process used by pathologists who assess cell neoplasticity and tumor extent at different microscope magnifications. The ratio of tumor cells can then be readily obtained by counting the number of cells in each class. To analyze an entire slide, we split it into multiple tiles that can be processed in parallel. The overall tumor cell ratio can then be aggregated. We perform experiments on a dataset of 100 slides with lung tumor specimens from both resection and tissue micro-array (TMA). We train fully-convolutional models using heavy data augmentation and batch normalization. On an unseen test set, we obtain an average mean absolute error on predicting the tumor cell ratio of less than 6%, which is significantly better than the human average of 20% and is key in properly selecting tissue samples for recent genetic panel tests geared at prescribing targeted cancer drugs. We perform ablation studies to show the importance of training two models at different magnifications and to justify the choice of some parameters, such as the size of the receptive field.