Every year radiologists face an increasing demand of image-based diagnosis from pa- tients, and computer-aided diagnosis (CAD) systems seem like a promising way to alleviate their workload. In recent years, many authors have proposed deep learning models to gen- erate reports from medical images, but they mainly focus on improving Natural Language Processing (NLP) metrics, such as BLEU and CIDEr, which may not be suitable to mea- sure clinical correctness in the reports, as indicated by multiple authors. Additionally, most approaches are end-to-end black box models that are difficult or impossible to understand by a human, which would make it very hard to implement in a clinical scenario. In this thesis, we contest the state-of-the-art models and evaluations in the report gen- eration from chest X-rays task. We provide further evidence showing that traditional NLP metrics are not enough to evaluate this task, by showing their lack of robustness in mul- tiple cases. For example, we show NLP metrics are not able to discriminate sentences with opposite clinical meaning, and we show that slightly altering report wording from a model can increase its NLP performance while maintaining high clinical performance. We also propose a template-based report generation model that detects a set of abnormalities and verbalizes them via fixed sentences into a structured report. We benchmark our model in the IU X-ray and MIMIC-CXR datasets against naive baselines, deep learning-based models, and literature models, by employing the CheXpert labeler and NLP metrics. The proposed model is much simpler and inherently interpretable than other state-of-the-art methods, and achieves better results in medical correctness metrics, though worse in NLP. We conclude there is a need to improve the assessment methods in this research area, by analyzing the available data in detail, performing more extensive evaluations and involving expert physicians.
How to join?
The defense will be in spanish. If you want to attend, please send an email to Pablo (email@example.com) before the defense.