arXiv subjects

Denis Musinguzi

Publications and source records attributed to Denis Musinguzi.

3 recordsLinked to original sources

arXiv2026

Location-Aware Pretraining for Medical Difference Visual Question Answering

Differential medical VQA models compare multiple images to identify clinically meaningful changes and rely on vision encoders to capture fine-grained visual differences that reflect radiologists' comparative diagnostic workflows. However, vision encoders trained using standard contrastive or classification objectives often fail to capture the subtle variations needed to distinguish true disease progression from acquisition-related variability. To address this limitation, we introduce a location-aware pretraining framework that incorporates automatic referring expressions (AREF), grounded captioning (GCAP), and conditional automatic referring expressions (CAREF). These tasks promote the learning of fine-grained, spatially grounded visual representations. When integrated with a language model, our approach achieves state-of-the-art performance on medical difference VQA by accurately identifying and reasoning about clinically relevant changes in chest X-ray images.

cs.CV↗

arXiv2025

Benchmarking Automatic Speech Recognition Models for African Languages

Automatic speech recognition (ASR) for African languages remains constrained by limited labeled data and the lack of systematic guidance on model selection, data scaling, and decoding strategies. Large pre-trained systems such as Whisper, XLS-R, MMS, and W2v-BERT have expanded access to ASR technology, but their comparative behavior in African low-resource contexts has not been studied in a unified and systematic way. In this work, we benchmark four state-of-the-art ASR models across 13 African languages, fine-tuning them on progressively larger subsets of transcribed data ranging from 1 to 400 hours. Beyond reporting error rates, we provide new insights into why models behave differently under varying conditions. We show that MMS and W2v-BERT are more data efficient in very low-resource regimes, XLS-R scales more effectively as additional data becomes available, and Whisper demonstrates advantages in mid-resource conditions. We also analyze where external language model decoding yields improvements and identify cases where it plateaus or introduces additional errors, depending on the alignment between acoustic and text resources. By highlighting the interaction between pre-training coverage, model architecture, dataset domain, and resource availability, this study offers practical and insights into the design of ASR systems for underrepresented languages.

cs.CL↗

arXiv2025

PaliGemma-CXR: A Multi-task Multimodal Model for TB Chest X-ray Interpretation

Tuberculosis (TB) is a infectious global health challenge. Chest X-rays are a standard method for TB screening, yet many countries face a critical shortage of radiologists capable of interpreting these images. Machine learning offers an alternative, as it can automate tasks such as disease diagnosis, and report generation. However, traditional approaches rely on task-specific models, which cannot utilize the interdependence between tasks. Building a multi-task model capable of performing multiple tasks poses additional challenges such as scarcity of multimodal data, dataset imbalance, and negative transfer. To address these challenges, we propose PaliGemma-CXR, a multi-task multimodal model capable of performing TB diagnosis, object detection, segmentation, report generation, and VQA. Starting with a dataset of chest X-ray images annotated with TB diagnosis labels and segmentation masks, we curated a multimodal dataset to support additional tasks. By finetuning PaliGemma on this dataset and sampling data using ratios of the inverse of the size of task datasets, we achieved the following results across all tasks: 90.32% accuracy on TB diagnosis and 98.95% on close-ended VQA, 41.3 BLEU score on report generation, and a mAP of 19.4 and 16.0 on object detection and segmentation, respectively. These results demonstrate that PaliGemma-CXR effectively leverages the interdependence between multiple image interpretation tasks to enhance performance.

cs.CV↗