A novel method for detection of pancreatic Ductal Adenocarcinoma using explainable machine learning.

BACKGROUND AND OBJECTIVE: Pancreatic Ductal Adenocarcinoma (PDAC) is a form of pancreatic cancer that is one of the primary causes of cancer-related deaths globally, with less than 10 % of the five years survival rate. The prognosis of pancreatic cancer has remained poor in the last four decades, mainly due to the lack of early diagnostic mechanisms. This study proposes a novel method for detecting PDAC using explainable and supervised machine learning from Raman spectroscopic signals. METHODS: An insightful feature set consisting of statistical, peak, and extended empirical mode decomposition features is selected using the support vector machine recursive feature elimination method integrated with a correlation bias reduction. Explicable features successfully identified mutations in Kirsten rat sarcoma viral oncogene homolog (KRAS) and tumor suppressor protein53 (TP53) in the fingerprint region for the first time in the literature. PDAC and normal pancreas are classified using K-nearest neighbor, linear discriminant analysis, and support vector machine classifiers. RESULTS: This study achieved a classification accuracy of 98.5% using a nonlinear support vector machine. Our proposed method reduced test time by 28.5 % and saved 85.6 % memory utilization, which reduces complexity significantly and is more accurate than the state-of-the-art method. The generalization of the proposed method is assessed by fifteen-fold cross-validation, and its performance is evaluated using accuracy, specificity, sensitivity, and receiver operating characteristic curves. CONCLUSIONS: In this study, we proposed a method to detect and define the fingerprint region for PDAC using explainable machine learning. This simple, accurate, and efficient method for PDAC detection in mice could be generalized to examine human pancreatic cancer and provide a basis for precise chemotherapy for early cancer treatment.

R2U++: a multiscale recurrent residual U-Net with dense skip connections for medical image segmentation.

U-Net is a widely adopted neural network in the domain of medical image segmentation. Despite its quick embracement by the medical imaging community, its performance suffers on complicated datasets. The problem can be ascribed to its simple feature extracting blocks: encoder/decoder, and the semantic gap between encoder and decoder. Variants of U-Net (such as R2U-Net) have been proposed to address the problem of simple feature extracting blocks by making the network deeper, but it does not deal with the semantic gap problem. On the other hand, another variant UNET++ deals with the semantic gap problem by introducing dense skip connections but has simple feature extraction blocks. To overcome these issues, we propose a new U-Net based medical image segmentation architecture R2U++. In the proposed architecture, the adapted changes from vanilla U-Net are: (1) the plain convolutional backbone is replaced by a deeper recurrent residual convolution block. The increased field of view with these blocks aids in extracting crucial features for segmentation which is proven by improvement in the overall performance of the network. (2) The semantic gap between encoder and decoder is reduced by dense skip pathways. These pathways accumulate features coming from multiple scales and apply concatenation accordingly. The modified architecture has embedded multi-depth models, and an ensemble of outputs taken from varying depths improves the performance on foreground objects appearing at various scales in the images. The performance of R2U++ is evaluated on four distinct medical imaging modalities: electron microscopy, X-rays, fundus, and computed tomography. The average gain achieved in IoU score is 1.5 ± 0.37% and in dice score is 0.9 ± 0.33% over UNET++, whereas, 4.21 ± 2.72 in IoU and 3.47 ± 1.89 in dice score over R2U-Net across different medical imaging segmentation datasets.

A new generative adversarial network for medical images super resolution.

For medical image analysis, there is always an immense need for rich details in an image. Typically, the diagnosis will be served best if the fine details in the image are retained and the image is available in high resolution. In medical imaging, acquiring high-resolution images is challenging and costly as it requires sophisticated and expensive instruments, trained human resources, and often causes operation delays. Deep learning based super resolution techniques can help us to extract rich details from a low-resolution image acquired using the existing devices. In this paper, we propose a new Generative Adversarial Network (GAN) based architecture for medical images, which maps low-resolution medical images to high-resolution images. The proposed architecture is divided into three steps. In the first step, we use a multi-path architecture to extract shallow features on multiple scales instead of single scale. In the second step, we use a ResNet34 architecture to extract deep features and upscale the features map by a factor of two. In the third step, we extract features of the upscaled version of the image using a residual connection-based mini-CNN and again upscale the feature map by a factor of two. The progressive upscaling overcomes the limitation for previous methods in generating true colors. Finally, we use a reconstruction convolutional layer to map back the upscaled features to a high-resolution image. Our addition of an extra loss term helps in overcoming large errors, thus, generating more realistic and smooth images. We evaluate the proposed architecture on four different medical image modalities: (1) the DRIVE and STARE datasets of retinal fundoscopy images, (2) the BraTS dataset of brain MRI, (3) the ISIC skin cancer dataset of dermoscopy images, and (4) the CAMUS dataset of cardiac ultrasound images. The proposed architecture achieves superior accuracy compared to other state-of-the-art super-resolution architectures.