An Interactive Visualization for Feature Localization in Deep Neural Networks

An Intelligent Method for Predicting Pacific Oyster (Crassostrea gigas) Freshness Using Deep Learning Fused with Malondialdehyde and Total Sulfhydryl Groups Information

To achieve a non-destructive and rapid detection of oyster freshness, an intelligent method using deep learning fused with malondialdehyde (MDA) and total sulfhydryl groups (SH) information was proposed. In this study, an "MDA-SH-storage days" polynomial fitting model and oyster meat image dataset were first built. AleNet-MDA and AlxNet-SH classification models were then constructed to automatically identify and classify four levels of oyster meat images with overall accuracies of 92.72% and 94.06%, respectively. Next, the outputs of the two models were used as the inputs to "MDA-SH-storage days" model, which ultimately succeeded in predicting the corresponding MDA content, SH content and storage day for an oyster image within 0.03 ms. Furthermore, the interpretability of the two models for oyster meat image were also investigated by feature visualization and strongest activations techniques. Thus, this study brings new thoughts on oyster freshness prediction from the perspective of computer vision and artificial intelligence.

Deep Learning-Based Intelligent Apple Variety Classification System and Model Interpretability Analysis

In this study, series networks (AlexNet and VGG-19) and directed acyclic graph (DAG) networks (ResNet-18, ResNet-50, and ResNet-101) with transfer learning were employed to identify and classify 13 classes of apples from 7439 images. Two training datasets, model evaluation metrics, and three visualization methods were used to objectively assess, compare, and interpret five Convolutional Neural Network (CNN)-based models. The results show that the dataset configuration had a significant impact on the classification results, as all models achieved over 96.1% accuracy on dataset A (training-to-testing = 2.4:1.0) compared to 89.4-93.9% accuracy on dataset B (training-to-testing = 1.0:3.7). VGG-19 achieved the highest accuracy of 100.0% on dataset A and 93.9% on dataset B. Moreover, for networks of the same framework, the model size, accuracy, and training and testing times increased as the model depth (number of layers) increased. Furthermore, feature visualization, strongest activations, and local interpretable model-agnostic explanations techniques were used to show the understanding of apple images by different trained models, as well as to reveal how and why the models make classification decisions. These results improve the interpretability and credibility of CNN-based models, which provides guidance for future applications of deep learning methods in agriculture.

Decoding the molecular subtypes of breast cancer seen on multimodal ultrasound images using an assembled convolutional neural network model: A prospective and multicentre study

Background

Preoperative determination of breast cancer molecular subtypes facilitates individualized treatment plan-making and improves patient prognosis. We aimed to develop an assembled convolutional neural network (ACNN) model for the preoperative prediction of molecular subtypes using multimodal ultrasound (US) images.

Methods

This multicentre study prospectively evaluated a dataset of greyscale US, colour Doppler flow imaging (CDFI), and shear-wave elastography (SWE) images in 807 patients with 818 breast cancers from November 2016 to February 2021. The St. Gallen molecular subtypes of breast cancer were confirmed by postoperative immunohistochemical examination. The monomodal ACNN model based on greyscale US images, the dual-modal ACNN model based on greyscale US and CDFI images, and the multimodal ACNN model based on greyscale US and CDFI as well as SWE images were constructed in the training cohort. The performances of three ACNN models in predicting four- and five-classification molecular subtypes and identifying triple negative from non-triple negative subtypes were assessed and compared. The performance of the multimodal ACNN was also compared with preoperative core needle biopsy (CNB).

Finding

The performance of the multimodal ACNN model (macroaverage area under the curve [AUC]: 0.89–0.96) was superior to that of the dual-modal ACNN model (macroaverage AUC: 0.81–0.84) and the monomodal ACNN model (macroaverage AUC: 0.73–0.75) in predicting four-classification breast cancer molecular subtypes, which was also better than that of preoperative CNB (AUC: 0.89–0.99 vs. 0.67–0.82, p < 0.05). In addition, the multimodal ACNN model outperformed the other two ACNN models in predicting five-classification molecular subtypes (AUC: 0.87–0.94 vs. 0.78-0.81 vs. 0.71–0.78) and identifying triple negative from non-triple negative breast cancers (AUC: 0.934–0.970 vs. 0.688–0.830 vs. 0.536–0.650, p < 0.05). Moreover, the multimodal ACNN model obtained satisfactory prediction performance for both T1 and non-T1 lesions (AUC: 0.957–0.958 and 0.932–0.985).

Interpretation

The multimodal US-based ACNN model is a potential noninvasive decision-making method for the management of patients with breast cancer in clinical practice.

Funding

This work was supported in part by the National Natural Science Foundation of China (Grants 81725008 and 81927801), Shanghai Municipal Health Commission (Grants 2019LJ21 and SHSLCZDZK03502), and the Science and Technology Commission of Shanghai Municipality (Grants 19441903200, 19DZ2251100, and 21Y11910800).

A generic intelligent tomato classification system for practical applications using DenseNet-201 with transfer learning

A generic intelligent tomato classification system based on DenseNet-201 with transfer learning was proposed and the augmented training sets obtained by data augmentation methods were employed to train the model. The trained model achieved high classification accuracy on the images of different quality, even those containing high levels of noise. Also, the trained model could accurately and efficiently identify and classify a single tomato image with only 29 ms, indicating that the proposed model has great potential value in real-world applications. The feature visualization of the trained models shows their understanding of tomato images, i.e., the learned common and high-level features. The strongest activations of the trained models show that the correct or incorrect target recognition areas by a model during the classification process will affect its final classification accuracy. Based on this, the results obtained in this study could provide guidance and new ideas to improve the development of intelligent agriculture.