Improving chest X-ray report generation by leveraging warm starting.

Automatically generating a report from a patient's Chest X-rays (CXRs) is a promising solution to reducing clinical workload and improving patient care. However, current CXR report generators-which are predominantly encoder-to-decoder models-lack the diagnostic accuracy to be deployed in a clinical setting. To improve CXR report generation, we investigate warm starting the encoder and decoder with recent open-source computer vision and natural language processing checkpoints, such as the Vision Transformer (ViT) and PubMedBERT. To this end, each checkpoint is evaluated on the MIMIC-CXR and IU X-ray datasets. Our experimental investigation demonstrates that the Convolutional vision Transformer (CvT) ImageNet-21K and the Distilled Generative Pre-trained Transformer 2 (DistilGPT2) checkpoints are best for warm starting the encoder and decoder, respectively. Compared to the state-of-the-art (M2 Transformer Progressive), CvT2DistilGPT2 attained an improvement of 8.3% for CE F-1, 1.8% for BLEU-4, 1.6% for ROUGE-L, and 1.0% for METEOR. The reports generated by CvT2DistilGPT2 have a higher similarity to radiologist reports than previous approaches. This indicates that leveraging warm starting improves CXR report generation. Code and checkpoints for CvT2DistilGPT2 are available at https://github.com/aehrc/cvt2distilgpt2.

Automated COVID-19 diagnosis and prognosis with medical imaging and who is publishing: a systematic review.

OBJECTIVES:  To conduct a systematic survey of published techniques for automated diagnosis and prognosis of COVID-19 diseases using medical imaging, assessing the validity of reported performance and investigating the proposed clinical use-case. To conduct a scoping review into the authors publishing such work. METHODS:  The Scopus database was queried and studies were screened for article type, and minimum source normalized impact per paper and citations, before manual relevance assessment and a bias assessment derived from a subset of the Checklist for Artificial Intelligence in Medical Imaging (CLAIM). The number of failures of the full CLAIM was adopted as a surrogate for risk-of-bias. Methodological and performance measurements were collected from each technique. Each study was assessed by one author. Comparisons were evaluated for significance with a two-sided independent t-test. FINDINGS:  Of 1002 studies identified, 390 remained after screening and 81 after relevance and bias exclusion. The ratio of exclusion for bias was 71%, indicative of a high level of bias in the field. The mean number of CLAIM failures per study was 8.3 ± 3.9 [1,17] (mean ± standard deviation [min,max]). 58% of methods performed diagnosis versus 31% prognosis. Of the diagnostic methods, 38% differentiated COVID-19 from healthy controls. For diagnostic techniques, area under the receiver operating curve (AUC) = 0.924 ± 0.074 [0.810,0.991] and accuracy = 91.7% ± 6.4 [79.0,99.0]. For prognostic techniques, AUC = 0.836 ± 0.126 [0.605,0.980] and accuracy = 78.4% ± 9.4 [62.5,98.0]. CLAIM failures did not correlate with performance, providing confidence that the highest results were not driven by biased papers. Deep learning techniques reported higher AUC (p < 0.05) and accuracy (p < 0.05), but no difference in CLAIM failures was identified. INTERPRETATION:  A majority of papers focus on the less clinically impactful diagnosis task, contrasted with prognosis, with a significant portion performing a clinically unnecessary task of differentiating COVID-19 from healthy. Authors should consider the clinical scenario in which their work would be deployed when developing techniques. Nevertheless, studies report superb performance in a potentially impactful application. Future work is warranted in translating techniques into clinical tools.

On training targets for deep learning approaches to clean speech magnitude spectrum estimation.

Estimation of the clean speech short-time magnitude spectrum (MS) is key for speech enhancement and separation. Moreover, an automatic speech recognition (ASR) system that employs a front-end relies on clean speech MS estimation to remain robust. Training targets for deep learning approaches to clean speech MS estimation fall into three categories: computational auditory scene analysis (CASA), MS, and minimum mean square error (MMSE) estimator training targets. The choice of the training target can have a significant impact on speech enhancement/separation and robust ASR performance. Motivated by this, the training target that produces enhanced/separated speech at the highest quality and intelligibility and that which is best for an ASR front-end is found. Three different deep neural network (DNN) types and two datasets, which include real-world nonstationary and coloured noise sources at multiple signal-to-noise ratio (SNR) levels, were used for evaluation. Ten objective measures were employed, including the word error rate of the Deep Speech ASR system. It is found that training targets that estimate the a priori SNR for MMSE estimators produce the highest objective quality scores. Moreover, it is established that the gain of MMSE estimators and the ideal amplitude mask produce the highest objective intelligibility scores and are most suitable for an ASR front-end.

Spectral distortion level resulting in a just-noticeable difference between an a priori signal-to-noise ratio estimate and its instantaneous case.

Minimum mean-square error (MMSE) approaches to speech enhancement are widely used in the literature. The quality of enhanced speech produced by an MMSE approach is directly impacted by the accuracy of the employed a priori signal-to-noise ratio (SNR) estimator. In this paper, the a priori SNR estimate spectral distortion (SD) level that results in a just-noticeable difference (JND) in the perceived quality of MMSE approach enhanced speech is found. The JND SD level is indicative of the accuracy that an a priori SNR estimator must exceed to have no impact on the perceived quality of MMSE approach enhanced speech. To measure the JND SD level, listening tests are conducted across five SNR levels, five noise sources, and two MMSE approaches [the MMSE short-time spectral amplitude (MMSE-STSA) estimator and the Wiener filter]. A statistical analysis of the results indicates that the JND SD level increases with the SNR level, is higher for the MMSE-STSA estimator, and is not impacted by the type of background noise. Following the literature, a significant improvement in a priori SNR estimation accuracy is required to reach the JND SD level.