HRKHu-Loss: harmonic regularization Kullback Huber loss-based lung cancer detection and severity level classification using CT image.

purposeEarly detection and accurate classification of lung cancer severity are crucial for improving patient survival rates and guiding clinicians to provide effective treatment strategies. Current diagnostic approaches predominantly depend on invasive techniques and manual analysis, which are time-consuming and subject to manual faults. There is a need for a reliable method to diagnose lung cancer and assess its severity level. To overcome these challenges, an efficient model named Harmonic Regularization Kullback Huber loss (HRKHu-Loss) is designed for the detection of cancer and to classify its severity level utilizing CT images.

methodsThe first step is to acquire the CT image from the database and bilateral filter pre-process the image. Moreover, segmentation is performed by the Region-based Online Selective Examination (ROSE) model. Then, features that include Patterns of Oriented Edge Magnitudes (POEM) with homogeneity as well as entropy, along with SqueezeNet, are extracted during feature extraction. Lung cancer is diagnosed by HRKHu-Loss, which is a loss function that modifies the layers of the Search Binary Neural Network (SBNN) and a Convolutional Neural Network (CNN). This integrated HRKHu-Loss function is the combination of the Adaptive Regularization term, Kullback-Leibler divergence loss and Huber loss. The model classifies CT images as either Normal or Abnormal. If an image is identified as Abnormal, the model further categorizes severity levels into mild, moderate or severe. Severity level classification is performed using the proposed HRKHu-Loss.

resultsThe devised HRKHu-Loss has gained superior values of accuracy, sensitivity, specificity, Matthews Correlation Coefficient (MCC), False Omission Rate (FOR) and Balanced Accuracy (BA) with values of 92.07%, 92.48%, 92.57%, 89.89%, 0.074 and 92.53%.

conclusionThe proposed HRKHu-Loss model demonstrates effective lung cancer detection and severity classification from CT images, highlighting its potential as a reliable diagnostic tool for clinical decision support.