Breast cancer diagnosis is one of the many areas that has taken advantage of artificial intelligence to achieve better performance, despite the fact that the availability of a large medical image dataset remains a challenge. Transfer learning (TL) is a phenomenon that enables deep learning algorithms to overcome the issue of shortage of training data in constructing an efficient model by transferring knowledge from a given source task to a target task. However, in most cases, ImageNet (natural images) pre-trained models that do not include medical images, are utilized for transfer learning to medical images. Considering the utilization of microscopic cancer cell line images that can be acquired in large amount, we argue that learning from both natural and medical datasets improves performance in ultrasound breast cancer image classification. The proposed multistage transfer learning (MSTL) algorithm was implemented using three pre-trained models: EfficientNetB2, InceptionV3, and ResNet50 with three optimizers: Adam, Adagrad, and stochastic gradient de-scent (SGD). Dataset sizes of 20,400 cancer cell images, 200 ultrasound images from Mendeley and 400 ultrasound images from the MT-Small-Dataset were used. ResNet50-Adagrad-based MSTL achieved a test accuracy of 99 ± 0.612% on the Mendeley dataset and 98.7 ± 1.1% on the MT-Small-Dataset, averaging over 5-fold cross validation. A p-value of 0.01191 was achieved when comparing MSTL against ImageNet based TL for the Mendeley dataset. The result is a significant improvement in the performance of artificial intelligence methods for ultrasound breast cancer classification compared to state-of-the-art methods and could remarkably improve the early diagnosis of breast cancer in young women.

Despite great achievements in classifying mammographic breast-mass images via deep-learning (DL), obtaining large amounts of training data and ensuring generalizations across different datasets with robust and well-optimized algorithms remain a challenge. ImageNet-based transfer learning (TL) and patch classifiers have been utilized to address these challenges. However, researchers have been unable to achieve the desired performance for DL to be used as a standalone tool. In this study, we propose a novel multi-stage TL from ImageNet and cancer cell line image pre-trained models to classify mammographic breast masses as either benign or malignant. We trained our model on three public datasets: Digital Database for Screening Mammography (DDSM), INbreast, and Mammographic Image Analysis Society (MIAS). In addition, a mixed dataset of the images from these three datasets was used to train the model. We obtained an average five-fold cross validation AUC of 1, 0.9994, 0.9993, and 0.9998 for DDSM, INbreast, MIAS, and mixed datasets, respectively. Moreover, the observed performance improvement using our method against the patch-based method was statistically significant, with a p-value of 0.0029. Furthermore, our patchless approach performed better than patch- and whole image-based methods, improving test accuracy by 8% (91.41% vs. 99.34%), tested on the INbreast dataset. The proposed method is of significant importance in solving the need for a large training dataset as well as reducing the computational burden in training and implementing the mammography-based deep-learning models for early diagnosis of breast cancer.

Microscopic image-based analysis has been intensively performed for pathological studies and diagnosis of diseases. However, mis-authentication of cell lines due to misjudgments by pathologists has been recognized as a serious problem. To address this problem, we propose a deep-learning-based approach for the automatic taxonomy of cancer cell types. A total of 889 bright-field microscopic images of four cancer cell lines were acquired using a benchtop microscope. Individual cells were further segmented and augmented to increase the image dataset. Afterward, deep transfer learning was adopted to accelerate the classification of cancer types. Experiments revealed that the deep-learning-based methods outperformed traditional machine-learning-based methods. Moreover, the Wilcoxon signed-rank test showed that deep ensemble approaches outperformed individual deep-learning-based models (p < 0.001) and were in effect to achieve the classification accuracy up to 97.735%. Additional investigation with the Wilcoxon signed-rank test was conducted to consider various network design choices, such as the type of optimizer, type of learning rate scheduler, degree of fine-tuning, and use of data augmentation. Finally, it was found that the using data augmentation and updating all the weights of a network during fine-tuning improve the overall performance of individual convolutional neural network models.

Point clouds and RGB images are naturally complementary modalities for 3D visual understanding - the former provides sparse but accurate locations of points on objects, while the latter contains dense color and texture information. Despite this potential for close sensor fusion, many methods train two models in isolation and use simple feature concatenation to represent 3D sensor data. This separated training scheme results in potentially sub-optimal performance and prevents 3D tasks from being used to benefit 2D tasks that are often useful on their own. To provide a more integrated approach, we propose a novel Multi-Modality Task Cascade network (MTC-RCNN) that leverages 3D box proposals to improve 2D segmentation predictions, which are then used to further refine the 3D boxes. We show that including a 2D network between two stages of 3D modules significantly improves both 2D and 3D task performance. Moreover, to prevent the 3D module from over-relying on the overfitted 2D predictions, we propose a dual-head 2D segmentation training and inference scheme, allowing the 2nd 3D module to learn to interpret imperfect 2D segmentation predictions. Evaluating our model on the challenging SUN RGB-D dataset, we improve upon state-of-the-art results of both single modality and fusion networks by a large margin ($\textbf{+3.8}$ mAP@0.5). Code will be released $\href{https://github.com/Divadi/MTC_RCNN}{\text{here.}}$

While numerous 3D detection works leverage the complementary relationship between RGB images and point clouds, developments in the broader framework of semi-supervised object recognition remain uninfluenced by multi-modal fusion. Current methods develop independent pipelines for 2D and 3D semi-supervised learning despite the availability of paired image and point cloud frames. Observing that the distinct characteristics of each sensor cause them to be biased towards detecting different objects, we propose DetMatch, a flexible framework for joint semi-supervised learning on 2D and 3D modalities. By identifying objects detected in both sensors, our pipeline generates a cleaner, more robust set of pseudo-labels that both demonstrates stronger performance and stymies single-modality error propagation. Further, we leverage the richer semantics of RGB images to rectify incorrect 3D class predictions and improve localization of 3D boxes. Evaluating on the challenging KITTI and Waymo datasets, we improve upon strong semi-supervised learning methods and observe higher quality pseudo-labels. Code will be released at https://github.com/Divadi/DetMatch