Convolutional neural network-based classification of cervical intraepithelial neoplasias using colposcopic image segmentation for acetowhite epithelium.

Colposcopy is a test performed to detect precancerous lesions of cervical cancer. Since cervical cancer progresses slowly, finding and treating precancerous lesions helps prevent cervical cancer. In particular, it is clinically important to detect high-grade squamous intraepithelial lesions (HSIL) that require surgical treatment among precancerous lesions of cervix. There have been several studies using convolutional neural network (CNN) for classifying colposcopic images. However, no studies have been reported on using the segmentation technique to detect HSIL. In present study, we aimed to examine whether the accuracy of a CNN model in detecting HSIL from colposcopic images can be improved when segmentation information for acetowhite epithelium is added. Without segmentation information, ResNet-18, 50, and 101 achieved classification accuracies of 70.2%, 66.2%, and 69.3%, respectively. The experts classified the same test set with accuracies of 74.6% and 73.0%. After adding segmentation information of acetowhite epithelium to the original images, the classification accuracies of ResNet-18, 50, and 101 improved to 74.8%, 76.3%, and 74.8%, respectively. We demonstrated that the HSIL detection accuracy improved by adding segmentation information to the CNN model, and the improvement in accuracy was consistent across different ResNets.

Enabling 100C Fast-Charging Bulk Bi Anodes for Na-Ion Batteries.

It is challenging to develop alloying anodes with ultrafast charging and large energy storage using bulk anode materials because of the difficulty of carrier-ion diffusion and fragmentation of the active electrode material. Herein, a rational strategy is reported to design bulk Bi anodes for Na-ion batteries that feature ultrafast charging, long cyclability, and large energy storage without using expensive nanomaterials and surface modifications. It is found that bulk Bi particles gradually transform into a porous nanostructure during cycling in a glyme-based electrolyte, whereas the resultant structure stores Na ions by forming phases with high Na diffusivity. These features allow the anodes to exhibit unprecedented electrochemical properties; the developed Na-Bi half-cell delivers 379 mA h g-1 (97% of that measured at 1C) at 7.7 A g-1 (20C) during 3500 cycles. It also retained 94% and 93% of the capacity measured at 1C even at extremely fast-charging rates of 80C and 100C, respectively. The structural origins of the measured properties are verified by experiments and first-principles calculations. The findings of this study not only broaden understanding of the underlying mechanisms of fast-charging anodes, but also provide basic guidelines for searching battery anodes that simultaneously exhibit high capacities, fast kinetics, and long cycling stabilities.