![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8171978%2F2bf2ecaf105d3d4d3fa1a96faa16294a%2F2567a20b32031caa71592fee2d9ccae.png?generation=1728699581616930&alt=media\")
First of all, I would like to sincerely thank the organizers for hosting this competition. I have learned a lot from it.
\nMy approach is very simple and follows a two-stage method. In the first stage, a point detection model is used to extract the ROI based on the predicted points. In the second stage, classification is performed on the ROI.
\nFirst stage
\nIn fact, I used a separate point detection model for each type of disease. It is important to note that for subarticular conditions, i can locate the corresponding vertebrae from the spinal points. Therefore, the point model for subarticular only involves localization on the x and y axes, which helps differentiate between left and right during the subsequent classification.
\n
Second stage
\nSince the first stage involves differentiating between left and right, there is no need for the model to perform this distinction in the second stage. I trained a separate classification model for each disease (2.5D CNN + GRU + AttentionHead), and developed a combined multi-modal model for the three diseases (32.5D CNN + GRU + 3AttentionHead). The final results are derived from the weighted average of these models. It is noteworthy that, in order to simulate potential inaccuracies in point localization from the first stage, I introduced random jitter within a specified range during the training of the second stage. This approach allowed the ROI to shift slightly within a confined area, thereby enhancing my performance. Furthermore, to address the issue of label distribution imbalance, I implemented oversampling of the data from the severe category. I used ResNet50 and SE-ResNeXt50 as the backbones for my models.
\nPost-process
\nI multiplied all the probabilities by a temperature value of 1.3.
\nFinally, I reviewed the approaches shared by everyone, and I found that some of my methods are also included among them. I won't elaborate further here. Thank you all, and I look forward to seeing you in the next competition!
", "messages": [ { "id": "3015102", "postDate": "10/12/2024 02:21:48", "content": "First of all, I would like to sincerely thank the organizers for hosting this competition. I have learned a lot from it.
\nMy approach is very simple and follows a two-stage method. In the first stage, a point detection model is used to extract the ROI based on the predicted points. In the second stage, classification is performed on the ROI.
\nFirst stage
\nIn fact, I used a separate point detection model for each type of disease. It is important to note that for subarticular conditions, i can locate the corresponding vertebrae from the spinal points. Therefore, the point model for subarticular only involves localization on the x and y axes, which helps differentiate between left and right during the subsequent classification.
\n
Second stage
\nSince the first stage involves differentiating between left and right, there is no need for the model to perform this distinction in the second stage. I trained a separate classification model for each disease (2.5D CNN + GRU + AttentionHead), and developed a combined multi-modal model for the three diseases (32.5D CNN + GRU + 3AttentionHead). The final results are derived from the weighted average of these models. It is noteworthy that, in order to simulate potential inaccuracies in point localization from the first stage, I introduced random jitter within a specified range during the training of the second stage. This approach allowed the ROI to shift slightly within a confined area, thereby enhancing my performance. Furthermore, to address the issue of label distribution imbalance, I implemented oversampling of the data from the severe category. I used ResNet50 and SE-ResNeXt50 as the backbones for my models.
\nPost-process
\nI multiplied all the probabilities by a temperature value of 1.3.
\nFinally, I reviewed the approaches shared by everyone, and I found that some of my methods are also included among them. I won't elaborate further here. Thank you all, and I look forward to seeing you in the next competition!
", "rawMarkdown": "First of all, I would like to sincerely thank the organizers for hosting this competition. I have learned a lot from it. \n\nMy approach is very simple and follows a two-stage method. In the first stage, a point detection model is used to extract the ROI based on the predicted points. In the second stage, classification is performed on the ROI.\n\n**First stage**\n\nIn fact, I used a separate point detection model for each type of disease. It is important to note that for subarticular conditions, i can locate the corresponding vertebrae from the spinal points. Therefore, the point model for subarticular only involves **localization on the x and y axes**, which helps differentiate between left and right during the subsequent classification.\n\n\n\n**Second stage**\n\nSince the first stage involves differentiating between left and right, there is no need for the model to perform this distinction in the second stage. I trained a separate classification model for each disease (2.5D CNN + GRU + AttentionHead), and developed a combined multi-modal model for the three diseases (3*2.5D CNN + GRU + 3*AttentionHead). The final results are derived from the weighted average of these models. It is noteworthy that, in order to simulate potential inaccuracies in point localization from the first stage, I introduced **random jitter** within a specified range during the training of the second stage. This approach allowed the ROI to shift slightly within a confined area, thereby enhancing my performance. Furthermore, to address the issue of label distribution imbalance, I implemented oversampling of the data from the severe category. I used ResNet50 and SE-ResNeXt50 as the backbones for my models.\n\n**Post-process**\n\nI multiplied all the probabilities by a temperature value of 1.3.\n\nFinally, I reviewed the approaches shared by everyone, and I found that some of my methods are also included among them. I won't elaborate further here. Thank you all, and I look forward to seeing you in the next competition!", "votes": null }, { "id": "3015109", "postDate": "10/12/2024 02:28:45", "content": "CV
\nss score: \nforaminal score: \nscs score: \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8171978%2F2070a13f48982b8b15e44f432da82359%2F32aa3ddef69c9a897c131651729e302.png?generation=1728701963282556&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8171978%2F5fd69340acfc7dcf8711dbd8c391f8c5%2F0efc7a6dc3b26cdafdf00d141805d99.png?generation=1728701975611840&alt=media\")
CV
\nss score: \nforaminal score: \nscs score: \n