Cancer Cell Classification Based on Morphological Features of 3D Phase Contrast Microscopy Using Deep Neural Network
Mi-Sun Kang, Jungyoon Kim · IEEE Access · 2025
3D phase contrast microscopy is one of the most common imaging modalities for the observation of long-term multicellular processes of living cells without phototoxicity and photobleaching, because the morphological features of cancer cells can be used as an indicator of metastasizing behavior. However, image features such as non-uniform illumination and phase contrast interference rings pose certain difficulties in analyzing these images. We propose a cancer cell classification methodology based on morphological features of 3D phase contrast microscopy and deep neural network with scaled principal component analysis. We initially apply non-uniform illumination correction based on the histogram information of images to correct unstable brightness problems in images and an image intensity-based global thresholding method to compensate for row-contrast artifacts via single-cell detection. We also extracted cross-sections to observe the morphological features using principal component analysis because of the nonsymmetric diffusion pattern of the interference that appeared around each cell. Then, the cell morphologies from an intensity gradient, considering local peaks as bright ring regions, were analyzed. The peak was calculated from the intensity profile from the center point of the cell area, which was the center of the extracted section, to the outer background. Based on the peak information, we extracted representative ten morphological features, applied a min-max scaler to convert the initial features, and used a deep neural network to classify active and inactive cancer cells. The proposed method achieved an area under the receiver operating characteristic curve value of 0.944 and an equal error rate of 0.091. We confirmed that the accuracy of classification using DNN with the proposed method was closer to the results of manual classification by experts, enabling a more precise analysis of cell morphology. This approach improves the accuracy of image-based cellular phenotypic profiling for assessing drug responses in patients.