Generación de informes radiológicos mediante el reconocimiento de voz

Abstract

The generation of radiology reports is a critical process that, in many departments, still relies on manual dictation and transcription or on generic ASR solutions poorly adapted to clinical Spanish. These approaches have recurrent drawbacks: high latency in report preparation, spelling and accentuation errors, terminological inconsistencies (acronyms and abbreviations), style variability, and, when cloud-based services are used, privacy concerns and dependency on connectivity. Dialectal variability, background noise, and lack of standardization further aggravate the problem, increasing the radiologist’s post-editing workload and reducing workflow efficiency. Therefore, a local, modular, and specialized solution is required—one that minimizes errors, preserves radiology records, and ensures data confidentiality. In this project, we implemented an on-premise automatic system for generating radiology reports in Spanish from dictated input, composed of five stages: (i) audio acquisition and acoustic preprocessing (normalization and resampling) to reduce input variability; (ii) automatic transcription using Whisper-small (a multilingual Transformer) configured for Spanish, with word-level timestamps to ensure traceability; (iii) lexical review assisted by a biomedical RoBERTa-based model aimed at normalizing clinical terms, diacritics, acronyms, and measurement units; (iv) stylistic adaptation using Gemma-2B-IT fine-tuned with LoRA/QLoRA (4-bit quantization for efficiency) to preserve the typical structure and tone of radiology reports without introducing undictated findings; and (v) postprocessing and report delivery through a graphical interface that displays intermediate versions (draft, corrected, and final) and facilitates expert post-editing. Evaluation was conducted on a set of 100 reports consisting of prompt–target pairs (error-prone dictation vs. expert-corrected version). Compared to the baseline configuration without specialization, the proposed system showed substantial improvements in automatic metrics: up to +112 % increase in BLEU and +77 % in ROUGE-2, along with consistent gains in ROUGE-1 and ROUGE-L. These results indicate reduced correction effort and better preservation of the narrative structure (finding → location → characteristics). The local architecture proved deployable on modest CPU/GPU setups, reducing reliance on external services and enhancing confidentiality. Overall, the proposal demonstrates both the technical feasibility and applied usefulness of the approach to accelerate report generation, improve terminological consistency, and reduce post-editing workload—laying a solid foundation for future extensions by subspecialty and the integration of clinical performance metrics.