Generalizability and robustness evaluation of attribute-based zero-shot learning.

In the field of deep learning, large quantities of data are typically required to effectively train models. This challenge has given rise to techniques like zero-shot learning (ZSL), which trains models on a set of "seen" classes and evaluates them on a set of "unseen" classes. Although ZSL has shown considerable potential, particularly with the employment of generative methods, its generalizability to real-world scenarios remains uncertain. The hypothesis of this work is that the performance of ZSL models is systematically influenced by the chosen "splits"; in particular, the statistical properties of the classes and attributes used in training. In this paper, we test this hypothesis by introducing the concepts of generalizability and robustness in attribute-based ZSL and carry out a variety of experiments to stress-test ZSL models against different splits. Our aim is to lay the groundwork for future research on ZSL models' generalizability, robustness, and practical applications. We evaluate the accuracy of state-of-the-art models on benchmark datasets and identify consistent trends in generalizability and robustness. We analyze how these properties vary based on the dataset type, differentiating between coarse- and fine-grained datasets, and our findings indicate significant room for improvement in both generalizability and robustness. Furthermore, our results demonstrate the effectiveness of dimensionality reduction techniques in improving the performance of state-of-the-art models in fine-grained datasets.

On the use of contrastive learning for standard-plane classification in fetal ultrasound imaging.

BACKGROUND: To investigate the effectiveness of contrastive learning, in particular SimClr, in reducing the need for large annotated ultrasound (US) image datasets for fetal standard plane identification. METHODS: We explore SimClr advantage in the cases of both low and high inter-class variability, considering at the same time how classification performance varies according to different amounts of labels used. This evaluation is performed by exploiting contrastive learning through different training strategies. We apply both quantitative and qualitative analyses, using standard metrics (F1-score, sensitivity, and precision), Class Activation Mapping (CAM), and t-Distributed Stochastic Neighbor Embedding (t-SNE). RESULTS: When dealing with high inter-class variability classification tasks, contrastive learning does not bring a significant advantage; whereas it results to be relevant for low inter-class variability classification, specifically when initialized with ImageNet weights. CONCLUSIONS: Contrastive learning approaches are typically used when a large number of unlabeled data is available, which is not representative of US datasets. We proved that SimClr either as pre-training with backbone initialized via ImageNet weights or used in an end-to-end dual-task may impact positively the performance over standard transfer learning approaches, under a scenario in which the dataset is small and characterized by low inter-class variability.

Classifying Vocal Folds Fixation from Endoscopic Videos with Machine Learning.

Vocal folds motility evaluation is paramount in both the assessment of functional deficits and in the accurate staging of neoplastic disease of the glottis. Diagnostic endoscopy, and in particular videoendoscopy, is nowadays the method through which the motility is estimated. The clinical diagnosis, however, relies on the examination of the videoendoscopic frames, which is a subjective and professional-dependent task. Hence, a more rigorous, objective, reliable, and repeatable method is needed. To support clinicians, this paper proposes a machine learning (ML) approach for vocal cords motility classification. From the endoscopic videos of 186 patients with both vocal cords preserved motility and fixation, a dataset of 558 images relative to the two classes was extracted. Successively, a number of features was retrieved from the images and used to train and test four well-grounded ML classifiers. From test results, the best performance was achieved using XGBoost, with precision = 0.82, recall = 0.82, F1 score = 0.82, and accuracy = 0.82. After comparing the most relevant ML models, we believe that this approach could provide precise and reliable support to clinical evaluation.Clinical Relevance- This research represents an important advancement in the state-of-the-art of computer-assisted otolaryngology, to develop an effective tool for motility assessment in the clinical practice.

Exploiting class activation mappings as prior to generate fetal brain ultrasound images with GANs.

The identification of fetal-head standard planes (FHSPs) from ultrasound (US) images is of fundamental importance to visualize cerebral structures and diagnose neural anomalies during gestation in a standardized way. To support the activity of healthcare operators, deep-learning algorithms have been proposed to classify these planes. To date, the translation of such algorithms in clinical practice is hampered by several factors, including the lack of large annotated datasets to train robust and generalizable algorithms. This paper proposes an approach to generate synthetic FHSP images with conditional generative adversarial network (cGAN), using class activation maps (CAMs) obtained from FHSP classification algorithms as cGAN conditional prior. Using the largest publicly available FHSP dataset, we generated realistic images of the three common FHSPs: trans-cerebellum, trans-thalamic and trans-ventricular. The evaluation through t-SNE shows the potential of the proposed approach to attenuate the problem of limited availability of annotated FHSP images.

A review on deep-learning algorithms for fetal ultrasound-image analysis.

Deep-learning (DL) algorithms are becoming the standard for processing ultrasound (US) fetal images. A number of survey papers in the field is today available, but most of them are focusing on a broader area of medical-image analysis or not covering all fetal US DL applications. This paper surveys the most recent work in the field, with a total of 153 research papers published after 2017. Papers are analyzed and commented from both the methodology and the application perspective. We categorized the papers into (i) fetal standard-plane detection, (ii) anatomical structure analysis and (iii) biometry parameter estimation. For each category, main limitations and open issues are presented. Summary tables are included to facilitate the comparison among the different approaches. In addition, emerging applications are also outlined. Publicly-available datasets and performance metrics commonly used to assess algorithm performance are summarized, too. This paper ends with a critical summary of the current state of the art on DL algorithms for fetal US image analysis and a discussion on current challenges that have to be tackled by researchers working in the field to translate the research methodology into actual clinical practice.

A deep learning approach to median nerve evaluation in ultrasound images of carpal tunnel inlet.

Ultrasound (US) imaging is recognized as a useful support for Carpal Tunnel Syndrome (CTS) assessment through the evaluation of median nerve morphology. However, US is still far to be systematically adopted to evaluate this common entrapment neuropathy, due to US intrinsic challenges, such as its operator dependency and the lack of standard protocols. To support sonographers, the present study proposes a fully-automatic deep learning approach to median nerve segmentation from US images. We collected and annotated a dataset of 246 images acquired in clinical practice involving 103 rheumatic patients, regardless of anatomical variants (bifid nerve, closed vessels). We developed a Mask R-CNN with two additional transposed layers at segmentation head to accurately segment the median nerve directly on transverse US images. We calculated the cross-sectional area (CSA) of the predicted median nerve. Proposed model achieved good performances both in median nerve detection and segmentation: Precision (Prec), Recall (Rec), Mean Average Precision (mAP) and Dice Similarity Coefficient (DSC) values are 0.916 ± 0.245, 0.938 ± 0.233, 0.936 ± 0.235 and 0.868 ± 0.201, respectively. The CSA values measured on true positive predictions were comparable with the sonographer manual measurements with a mean absolute error (MAE) of 0.918 mm2. Experimental results showed the potential of proposed model, which identified and segmented the median nerve section in normal anatomy images, while still struggling when dealing with infrequent anatomical variants. Future research will expand the dataset including a wider spectrum of normal anatomy and pathology to support sonographers in daily practice.

Development of a convolutional neural network for the identification and the measurement of the median nerve on ultrasound images acquired at carpal tunnel level.

BACKGROUND: Deep learning applied to ultrasound (US) can provide a feedback to the sonographer about the correct identification of scanned tissues and allows for faster and standardized measurements. The most frequently adopted parameter for US diagnosis of carpal tunnel syndrome is the increasing of the cross-sectional area (CSA) of the median nerve. Our aim was to develop a deep learning algorithm, relying on convolutional neural networks (CNNs), for the localization and segmentation of the median nerve and the automatic measurement of its CSA on US images acquired at the proximal inlet of the carpal tunnel. METHODS: Consecutive patients with rheumatic and musculoskeletal disorders were recruited. Transverse US images were acquired at the carpal tunnel inlet, and the CSA was manually measured. Anatomical variants were registered. The dataset consisted of 246 images (157 for training, 40 for validation, and 49 for testing) from 103 patients each associated with manual annotations of the nerve boundary. A Mask R-CNN, state-of-the-art CNN for image semantic segmentation, was trained on this dataset to accurately localize and segment the median nerve section. To evaluate the performances on the testing set, precision (Prec), recall (Rec), mean average precision (mAP), and Dice similarity coefficient (DSC) were computed. A sub-analysis excluding anatomical variants was performed. The CSA was automatically measured by the algorithm. RESULTS: The algorithm correctly identified the median nerve in 41/49 images (83.7%) and in 41/43 images (95.3%) excluding anatomical variants. The following metrics were obtained (with and without anatomical variants, respectively): Prec 0.86 ± 0.33 and 0.96 ± 0.18, Rec 0.88 ± 0.33 and 0.98 ± 0.15, mAP 0.88 ± 0.33 and 0.98 ± 0.15, and DSC 0.86 ± 0.19 and 0.88 ± 0.19. The agreement between the algorithm and the sonographer CSA measurements was excellent [ICC 0.97 (0.94-0.98)]. CONCLUSIONS: The developed algorithm has shown excellent performances, especially if excluding anatomical variants. Future research should aim at expanding the US image dataset including a wider spectrum of normal anatomy and pathology. This deep learning approach has shown very high potentiality for a fully automatic support for US assessment of carpal tunnel syndrome.

A deep-learning framework for metacarpal-head cartilage-thickness estimation in ultrasound rheumatological images.

OBJECTIVE: Rheumatoid arthritis (RA) is a chronic disease characterized by erosive symmetrical polyarthritis. Bone and cartilage are the main joint targets of this disease. Cartilage damage is one of the most relevant determinants of physical disability in RA patients. Cartilage damage is nowadays assessed by clinicians, which manually measure cartilage thickness in ultrasound (US) imaging. This poses issues relevant to intra-and inter-observer variability. Relying on the acquisition of metacarpal-head US images from 38 subjects, this work addresses the problem of automatic cartilage-thickness measurement by designing a new deep-learning (DL) framework. METHODS: The framework consists of a Convolutional Neural Network (CNN), responsible for regressing cartilage-interface distance fields, followed by a post-processing step to delineate the two cartilage interfaces from the predicted distance fields and compute the cartilage thickness. RESULTS: Our framework achieved encouraging results with a mean absolute difference (ADF) of 0.032 (±0.026) mm against manual thickness annotation by an expert clinician. When evaluating the intra-observer variability, we obtained an ADF = 0.036 (±0.028) mm. CONCLUSION: The proposed framework achieved an ADF against manual annotation that was comparable to the intra-observer variability, proving the potential of DL in the field. SIGNIFICANCE: This work is the first to address the problem of automatic cartilage-thickness estimation in US rheumatological images with DL, paving the way for future research in the field. From a clinical perspective, the proposed framework proved to be a valuable tool to support the clinical routine increasing the reproducibility of cartilage thickness measurements.

Mask-R[Formula: see text]CNN: a distance-field regression version of Mask-RCNN for fetal-head delineation in ultrasound images.

BACKGROUND AND OBJECTIVES: Fetal head-circumference (HC) measurement from ultrasound (US) images provides useful hints for assessing fetal growth. Such measurement is performed manually during the actual clinical practice, posing issues relevant to intra- and inter-clinician variability. This work presents a fully automatic, deep-learning-based approach to HC delineation, which we named Mask-R[Formula: see text]CNN. It advances our previous work in the field and performs HC distance-field regression in an end-to-end fashion, without requiring a priori HC localization nor any postprocessing for outlier removal. METHODS: Mask-R[Formula: see text]CNN follows the Mask-RCNN architecture, with a backbone inspired by feature-pyramid networks, a region-proposal network and the ROI align. The Mask-RCNN segmentation head is here modified to regress the HC distance field. RESULTS: Mask-R[Formula: see text]CNN was tested on the HC18 Challenge dataset, which consists of 999 training and 335 testing images. With a comprehensive ablation study, we showed that Mask-R[Formula: see text]CNN achieved a mean absolute difference of 1.95 mm (standard deviation [Formula: see text] mm), outperforming other approaches in the literature. CONCLUSIONS: With this work, we proposed an end-to-end model for HC distance-field regression. With our experimental results, we showed that Mask-R[Formula: see text]CNN may be an effective support for clinicians for assessing fetal growth.

Artificial Intelligence for Ultrasound Informative Image Selection of Metacarpal Head Cartilage. A Pilot Study.

Objectives: This study aims to develop an automatic deep-learning algorithm, which is based on Convolutional Neural Networks (CNNs), for ultrasound informative-image selection of hyaline cartilage at metacarpal head level. The algorithm performance and that of three beginner sonographers were compared with an expert assessment, which was considered the gold standard. Methods: The study was divided into two steps. In the first one, an automatic deep-learning algorithm for image selection was developed using 1,600 ultrasound (US) images of the metacarpal head cartilage (MHC) acquired in 40 healthy subjects using a very high-frequency probe (up to 22 MHz). The algorithm task was to identify US images defined informative as they show enough information to fulfill the Outcome Measure in Rheumatology US definition of healthy hyaline cartilage. The algorithm relied on VGG16 CNN, which was fine-tuned to classify US images in informative and non-informative ones. A repeated leave-four-subject out cross-validation was performed using the expert sonographer assessment as gold-standard. In the second step, the expert assessed the algorithm and the beginner sonographers' ability to obtain US informative images of the MHC. Results: The VGG16 CNN showed excellent performance in the first step, with a mean area (AUC) under the receiver operating characteristic curve, computed among the 10 models obtained from cross-validation, of 0.99 ± 0.01. The model that reached the best AUC on the testing set, which we named "MHC identifier 1," was then evaluated by the expert sonographer. The agreement between the algorithm, and the expert sonographer was almost perfect [Cohen's kappa: 0.84 (95% confidence interval: 0.71-0.98)], whereas the agreement between the expert and the beginner sonographers using conventional assessment was moderate [Cohen's kappa: 0.63 (95% confidence interval: 0.49-0.76)]. The conventional obtainment of US images by beginner sonographers required 6.0 ± 1.0 min, whereas US videoclip acquisition by a beginner sonographer lasted only 2.0 ± 0.8 min. Conclusion: This study paves the way for the automatic identification of informative US images for assessing MHC. This may redefine the US reliability in the evaluation of MHC integrity, especially in terms of intrareader reliability and may support beginner sonographers during US training.

A regression framework to head-circumference delineation from US fetal images.

BACKGROUND AND OBJECTIVES: Measuring head-circumference (HC) length from ultrasound (US) images is a crucial clinical task to assess fetus growth. To lower intra- and inter-operator variability in HC length measuring, several computer-assisted solutions have been proposed in the years. Recently, a large number of deep-learning approaches is addressing the problem of HC delineation through the segmentation of the whole fetal head via convolutional neural networks (CNNs). Since the task is a edge-delineation problem, we propose a different strategy based on regression CNNs. METHODS: The proposed framework consists of a region-proposal CNN for head localization and centering, and a regression CNN for accurately delineate the HC. The first CNN is trained exploiting transfer learning, while we propose a training strategy for the regression CNN based on distance fields. RESULTS: The framework was tested on the HC18 Challenge dataset, which consists of 999 training and 335 testing images. A mean absolute difference of 1.90 ( ±  1.76) mm and a Dice similarity coefficient of 97.75 ( ±  1.32) % were achieved, overcoming approaches in the literature. CONCLUSIONS: The experimental results showed the effectiveness of the proposed framework, proving its potential in supporting clinicians during the clinical practice.