Applying Deep Learning Techniques in Automated Analysis of CT scan images for the Detection and Localization of Renal Cell Carcinoma.pptx
- 1. Applying Deep Learning Techniques in Automated Analysis of CT scan images for the Detection and Localization of Renal Cell Carcinoma
- 2. What is Renal Cell Carcinoma?
- 3. Risk factors of Renal Cell Carcinoma Smoking Obesity Hypertension Occupational exposure (metal, chemical, rubber, and printing industries) Cadmium exposure Radiation therapy Dietary (high fat/protein and low fruits/vegetables)
- 4. Symptoms and Treatment • Flank Pain • Hematuria • Palpable Mass
- 5. Related work Koo, J. C., Hum, Y. C., Lai, K. W., Yap, W. S., Manickam, S., & Tee, Y. K. (2022, March). Deep Machine Learning Histopathological Image Analysis for Renal Cancer Detection. In Proceedings of the 8th International Conference on Computing and Artificial Intelligence (pp. 657-663). Xu, L., Yang, C., Zhang, F., Cheng, X., Wei, Y., Fan, S., ... & Song, B. (2022). Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers, 14(11), 2574. Gharaibeh, M., Alzu’bi, D., Abdullah, M., Hmeidi, I., Al Nasar, M. R., Abualigah, L., & Gandomi, A. H. (2022). Radiology imaging scans for early diagnosis of kidney tumors: a review of data analytics-based machine learning and deep learning approaches. Big Data and Cognitive Computing, 6(1), 29. Uhm, K. H., Jung, S. W., Choi, M. H., Shin, H. K., Yoo, J. I., Oh, S. W., ... & Ko, S. J. (2021). Deep learning for end-to-end kidney cancer diagnosis on multi-phase abdominal computed tomography. NPJ Precision Oncology, 5(1), 1-6. Nikpanah, M., Xu, Z., Jin, D., Farhadi, F., Saboury, B., Ball, M. W., ... & Malayeri, A. A. (2021). A deep-learning based artificial intelligence (AI) approach for differentiation of clear cell renal cell carcinoma from oncocytoma on multi-phasic MRI. Clinical Imaging, 77, 291-298. Lee, S., Jung, J., Park, I., Park, K., & Kim, D. S. (2020). A deep learning and similarity-based hierarchical clustering approach for pathological stage prediction of papillary renal cell carcinoma. Computational and structural biotechnology journal, 18, 2639-2646. Shon, H. S., Batbaatar, E., Kim, K. O., Cha, E. J., & Kim, K. A. (2020). Classification of kidney cancer data using cost-sensitive hybrid deep learning approach. Symmetry, 12(1), 154. Fuat, T. Ü. R. K., Murat, L. Ü. Y., & BARIŞÇI, N. (2019). Machine Learning of Kidney Tumors and Diagnosis and Classification by Deep Learning Methods. International Journal of Engineering Research and Development, 11(3), 802-812. Kaur, G., Kalia, G., & Sondhi, P. (2019). Artificial Neural Network Based Detection of Renal Tumors using CT Scan Image Processing. Tuncer, S. A., & Alkan, A. (2018). A decision support system for detection of the renal cell cancer in the kidney. Measurement, 123, 298-303.
- 6. Deep learning for end-to-end kidney cancer diagnosis on multi-phase abdominal computed tomography Limitations of the previous work There are strong overlaps in image-level features between renal tumor subtypes, which make subtype classification difficult and cause inter-observer variation. In most prior studies on tumor classification, lesions were classified into only two classes (benign/malignant) or the three RCC classes (ccRCC, pRCC, and chRCC). The previous diagnosis systems required the manual lesion identification process, in which the regions of tumors are drawn by radiologists.
- 7. Database A large dataset consisting of 1035 CT images from 308 patients who underwent nephrectomy for renal tumors between 2003 and 2020. This dataset contains five major subtypes of renal tumors including both benign and malignant tumors: oncocytoma, AML, chRCC, pRCC, and ccRCC. Table presenting Patient demographics, subtype, and tumor size distributions for training/test dataset.
- 8. Methodology • The framework(Figure 1) takes a multi-phase CT scan as an input. • The framework first produces the kidney and tumor masks for each phase using a shared 3D segmentation model. • The framework then aligns the tumor regions across phases and outputs a probability distribution over five subtype classes of renal tumor through a classification model. • In the segmentation results, the green and magenta represent the kidney and tumor, respectively.
- 9. The framework first extracts the kidney and tumor masks from the whole CT volume for each phase using the three-dimensional (3D) CNN-based segmentation model and voxel-level segmentation labels were used to train this model. Then, the CT volumes of different phases are aligned based on the segmented regions, and finally, the CNN-based classification model analyzes the aligned tumor regions and predicts the subtype. Postoperative pathology- confirmed tumor labels were used to train the classification model. For each patient, multiple phases were acquired at different times such as non-contrast, arterial (20–30s after contrast injection), portal (60–70s), and delayed (>180 s) phases. Voxel-level segmentation labels are obtained for each CT scan, where trained annotators manually delineated kidneys and tumors in the images and then a radiologist (experience of 11 years) refined the annotations.
- 10. Model development The proposed model has three main components: kidney and tumor segmentation, multi-phase alignment, and tumor subtype classification. 3D U-Net(trained on 848 CT scans including four different contrast phases) used for kidney and tumor segmentation, where the network classifies each voxel in a CT volume into three classes: background, kidney, and tumor. The CT volumes were resampled to a 1.5 × 1.5 × 3 mm3 voxel size. The network parameters were then optimized using stochastic gradient descent on the sum of the cross-entropy and Dice loss function. The hyperparameters required for training, such as the batch size and learning rate, were chosen by nnU-Net. The transformation parameters were iteratively updated to align the kidney and tumor masks of the two phases until convergence. Dice loss was minimized using an Adam optimizer with a learning rate of 0.01. ResNet-101 is used to classify the pathological subtypes of renal tumor. For each case, the slice with the largest segmented tumor area was extracted from each phase of the CT scans, and the rectangular region containing the tumor region was then cropped from each extracted slice. The cropped images were then resized to 224 × 224 pixels and concatenated to form a 3-channel image, which was used as the input to the classification network.
- 11. In the testing stage, the results of the network was averaged from three 3-channel images. ResNet-101 was initialized with the weights pre- trained on ImageNet ,and added a 1 × 1 × 1 convolutional layer at the beginning of the network and changed the last fully connected layer to produce a distribution over five classes. The network was trained using the cross-entropy loss with stochastic gradient descent. The final component outputs the probability for each subtype class. Performance of segmentation on the test dataset They evaluated the segmentation model by measuring the Dice similarity coefficient (DSC), which quantifies the volume overlap between manual annotations and the masks produced by the model for the kidney and tumor regions. The average DSCs for the kidney and tumor were obtained as 0.969±0.014 and 0.856± 0.131, respectively.
- 12. Performance Figure 2a shows the receiver operating characteristic (ROC) curves of the model and the performance of the radiologists. The area under the curves (AUCs) with 95% confidence interval (CI) for each curve. The model achieved an average AUC of 0.889 (95% CI, 0.827–0.945), and exceeded both the top-1 and top- 2 performance of the radiologists in most cases. In particular, the points indicating the average performance of the radiologists fell on or below the ROC curves of the model for all subtype classes.
- 13. Figure 2b shows the confusion matrices for the model and all individual radiologists. It was observed that chRCC, AML, and oncocytoma were frequently misclassified as ccRCC by the radiologists, whereas they were more correctly classified by the model. The model achieved the accuracy of 0.72, exceeding both the average top-1 and top-2 accuracy of radiologists, which were 0.42 and 0.56, respectively.
- 14. Compared to the average radiologist, the model demonstrated statistically significant improvements in top-1 sensitivity (P < 0.05) for chRCC and AML, and even in top-2 sensitivity (P < 0.05) for AML (Fig. 2c) and there is significant improvements in specificity (P<0.05) for ccRCC and oncocytoma (Fig. 2d).