Strengthening deep-learning models for intracranial hemorrhage detection: strongly annotated computed tomography images and model ensembles

Background and purpose

Multiple attempts at intracranial hemorrhage (ICH) detection using deep-learning techniques have been plagued by clinical failures. We aimed to compare the performance of a deep-learning algorithm for ICH detection trained on strongly and weakly annotated datasets, and to assess whether a weighted ensemble model that integrates separate models trained using datasets with different ICH improves performance.

Methods

We used brain CT scans from the Radiological Society of North America (27,861 CT scans, 3,528 ICHs) and AI-Hub (53,045 CT scans, 7,013 ICHs) for training. DenseNet121, InceptionResNetV2, MobileNetV2, and VGG19 were trained on strongly and weakly annotated datasets and compared using independent external test datasets. We then developed a weighted ensemble model combining separate models trained on all ICH, subdural hemorrhage (SDH), subarachnoid hemorrhage (SAH), and small-lesion ICH cases. The final weighted ensemble model was compared to four well-known deep-learning models. After external testing, six neurologists reviewed 91 ICH cases difficult for AI and humans.

Results

InceptionResNetV2, MobileNetV2, and VGG19 models outperformed when trained on strongly annotated datasets. A weighted ensemble model combining models trained on SDH, SAH, and small-lesion ICH had a higher AUC, compared with a model trained on all ICH cases only. This model outperformed four deep-learning models (AUC [95% C.I.]: Ensemble model, 0.953[0.938-0.965]; InceptionResNetV2, 0.852[0.828-0.873]; DenseNet121, 0.875[0.852-0.895]; VGG19, 0.796[0.770-0.821]; MobileNetV2, 0.650[0.620-0.680]; p < 0.0001). In addition, the case review showed that a better understanding and management of difficult cases may facilitate clinical use of ICH detection algorithms.

Conclusion

We propose a weighted ensemble model for ICH detection, trained on large-scale, strongly annotated CT scans, as no model can capture all aspects of complex tasks.

A Fully Flexible Intelligent Thermal Touch Panel Based on Intrinsically Plastic Ag2 S Semiconductor

Wearable touch panels, a typical flexible electronic device, can recognize and feed back the information of finger touch and movement. Excellent wearable touch panels are required to accurately and quickly monitor the signals of finger movement as well as the capacity of bearing various types of deformation. High-performance thermistor materials are one of the key functional components, but to date, a long-standing bottleneck is that inorganic semiconductors are typically brittle while the electrical properties of organic semiconductors are quite low. Herein, a high-performance flexible temperature sensor is reported by using plastic Ag₂ S with ultrahigh temperature coefficient of resistance of -4.7% K^-1 and resolution of 0.05 K, and rapid response/recovery time of 0.11/0.11 s. Moreover, the temperature sensor shows excellent durability without performance damage or loss during force stimuli tests. In addition, a fully flexible intelligent touch panel composed of a 16 × 10 Ag₂ S-film-based temperature sensor array, as well as a flexible printed circuit board and a deep-learning algorithm is designed for perceiving finger touch signals in real-time, and intelligent feedback of Chinese characters and letters on an app. These results strongly show that high-performance flexible inorganic semiconductors can be widely used in flexible electronics.

Development and validation of a deep-learning algorithm for the detection of neovascular age-related macular degeneration from colour fundus photographs

IMPORTANCE

Detection of early onset neovascular age-related macular degeneration (AMD) is critical to protecting vision.

BACKGROUND

To describe the development and validation of a deep-learning algorithm (DLA) for the detection of neovascular age-related macular degeneration.

DESIGN

Development and validation of a DLA using retrospective datasets.

PARTICIPANTS

We developed and trained the DLA using 56 113 retinal images and an additional 86 162 images from an independent dataset to externally validate the DLA. All images were non-stereoscopic and retrospectively collected.

METHODS

The internal validation dataset was derived from real-world clinical settings in China. Gold standard grading was assigned when consensus was reached by three individual ophthalmologists. The DLA classified 31 247 images as gradable and 24 866 as ungradable (poor quality or poor field definition). These ungradable images were used to create a classification model for image quality. Efficiency and diagnostic accuracy were tested using 86 162 images derived from the Melbourne Collaborative Cohort Study. Neovascular AMD and/or ungradable outcome in one or both eyes was considered referable.

MAIN OUTCOME MEASURES

Area under the receiver operating characteristic curve (AUC), sensitivity and specificity.

RESULTS

In the internal validation dataset, the AUC, sensitivity and specificity of the DLA for neovascular AMD was 0.995, 96.7%, 96.4%, respectively. Testing against the independent external dataset achieved an AUC, sensitivity and specificity of 0.967, 100% and 93.4%, respectively. More than 60% of false positive cases displayed other macular pathologies. Amongst the false negative cases (internal validation dataset only), over half (57.2%) proved to be undetected detachment of the neurosensory retina or RPE layer.

CONCLUSIONS AND RELEVANCE

This DLA shows robust performance for the detection of neovascular AMD amongst retinal images from a multi-ethnic sample and under different imaging protocols. Further research is warranted to investigate where this technology could be best utilized within screening and research settings.