Wafer map failure pattern classification using geometric transformation-invariant convolutional neural network.

Wafer map defect pattern classification is essential in semiconductor manufacturing processes for increasing production yield and quality by providing key root-cause information. However, manual diagnosis by field experts is difficult in large-scale production situations, and existing deep-learning frameworks require a large quantity of data for learning. To address this, we propose a novel rotation- and flip-invariant method based on the labeling rule that the wafer map defect pattern has no effect on the rotation and flip of labels, achieving class discriminant performance in scarce data situations. The method utilizes a convolutional neural network (CNN) backbone with a Radon transformation and kernel flip to achieve geometrical invariance. The Radon feature serves as a rotation-equivariant bridge for translation-invariant CNNs, while the kernel flip module enables the model to be flip-invariant. We validated our method through extensive qualitative and quantitative experiments. For qualitative analysis, we suggest a multi-branch layer-wise relevance propagation to properly explain the model decision. For quantitative analysis, the superiority of the proposed method was validated with an ablation study. In addition, we verified the generalization performance of the proposed method to rotation and flip invariants for out-of-distribution data using rotation and flip augmented test sets.

Deep learning predicts the differentiation of kidney organoids derived from human induced pluripotent stem cells.

BACKGROUND: Kidney organoids derived from human pluripotent stem cells (hPSCs) contain multilineage nephrogenic progenitor cells and can recapitulate the development of the kidney. Kidney organoids derived from hPSCs have the potential to be applied in regenerative medicine as well as renal disease modeling, drug screening, and nephrotoxicity testing. Despite biotechnological advances, individual differences in morphological and growth characteristics among kidney organoids need to be addressed before clinical and commercial application. In this study, we hypothesized that an automated noninvasive method based on deep learning of bright-field images of kidney organoids can predict their differentiation status. METHODS: Bright-field images of kidney organoids were collected on day 18 after differentiation. To train convolutional neural networks (CNNs), we utilized a transfer learning approach. CNNs were trained to predict the differentiation of kidney organoids on bright-field images based on the messenger RNA expression of renal tubular epithelial cells as well as podocytes. RESULTS: The best prediction model was DenseNet121 with a total Pearson correlation coefficient score of 0.783 on a test dataset. W classified the kidney organoids into two categories: organoids with above-average gene expression (Positive) and those with below-average gene expression (Negative). Comparing the best-performing CNN with human-based classifiers, the CNN algorithm had a receiver operating characteristic-area under the curve (AUC) score of 0.85, while the experts had an AUC score of 0.48. CONCLUSION: These results confirmed our original hypothesis and demonstrated that our artificial intelligence algorithm can successfully recognize the differentiation status of kidney organoids.

Predicting multipotency of human adult stem cells derived from various donors through deep learning.

Adult stem cell-based therapeutic approaches have great potential in regenerative medicine because of their immunoregulatory properties and multidifferentiation capacity. Nevertheless, the outcomes of stem cell­based therapies to date have shown inconsistent efficacy owing to donor variation, thwarting the expectation of clinical effects. However, such donor dependency has been elucidated by biological consequences that current research could not predict. Here, we introduce cellular morphology-based prediction to determine the multipotency rate of human nasal turbinate stem cells (hNTSCs), aiming to predict the differentiation rate of keratocyte progenitors. We characterized the overall genes and morphologies of hNTSCs from five donors and compared stemness-related properties, including multipotency and specific lineages, using mRNA sequencing. It was demonstrated that transformation factors affecting the principal components were highly related to cell morphology. We then performed a convolutional neural network-based analysis, which enabled us to assess the multipotency level of each cell group based on their morphologies with 85.98% accuracy. Surprisingly, the trend in expression levels after ex vivo differentiation matched well with the deep learning prediction. These results suggest that AI­assisted cellular behavioral prediction can be utilized to perform quantitative, non-invasive, single-cell, and multimarker characterizations of live stem cells for improved quality control in clinical cell therapies.