Er Yunjian A team specializing in scientific research
Welcome to like + favorite + follow
The current high-score student letter articles, in order to break through the involution, have turned to "pathological image analysis", "radiomics analysis", this direction and then use "machine learning, deep learning models", the score rises swish~~
Fans and friends of "Yunshengxin", have you seen the new ideas such as bioletter + pathological analysis and radiomics recommended by Xiaoyun? Friends who haven't seen it yet, hurry up and make up for it (ps: interested partners welcome to click the link at the end of the article to watch), this is a good opportunity to sprint high-scoring students' letters, we can't let it go directly!
Today, Xiaoyun will use an 11+ article to show it with an example! This paper uses public pathology image data, mutation data, and clinical variables to construct a deep learning model to stratify tumors. The main body of analysis lies in the construction and analysis of deep learning models, and the first innovation in analysis methods is very high. Secondly, the analysis target is public pathological images and mutations, which are also highly innovative data types compared to conventional transcriptome data, so comprehensively, the innovation of double-layer superposition is indeed leveraged, worthy of 11 points + LANCET sub-journal! Such a good idea of getting a high score for life letters, if you don't learn, you will lose! Without saying much, follow Xiaoyun to find out~ ~
Title: Integrating Deep Learning Analysis to Improve Death Risk Stratified in Patients with Colon Adenocarcinoma
Journal: EBioMedicine
Impact factor: IF=11.1
Published: July 2023
Research background
Colorectal cancer is the fourth most diagnosed cancer and the second most fatal. Many clinical variables, pathological features, and genomic features correlate with patient risk, but reliable patient stratification in the clinic remains a challenging task. Here, the authors evaluate how to combine pathology images, clinical and genomic features to predict risk.
Data sources
Research ideas
Using a dataset of 108 COAD patients from TCGA as a training set, a comprehensive deep learning model was developed that combines paraffin-embedded (FFPE) full-section images (WSIs), clinical variables, and mutational features to stratify patients according to their risk of death. Subsequently, the training set and WSU and TCGA-READ validation focused on evaluating the hierarchical effect of the model on mortality risk.
1) Data preprocessing: The tumor area of WSIs is annotated by 3 pathologist experts by multi-scale observation of the specimen. Five clinical variables related to patient outcomes were selected: age, sex, and TNM stage of colon adenocarcinoma (tumor (T), lymph node (N), and metastasis (M)). 207 genes from 11 typical cancer pathways and 11 genes with the most common mutations in TCGA-COAD were selected, and a 10% threshold was used to filter out genes that were not frequently mutated in TCGA-COAD patients, resulting in a total of 26 genes.
2) Model training: a: Pure image model, the author used the Inception V3 model pre-trained on the ImageNet database to build an image model based on WSIs. b: Integrate the model to connect Inception V3 model features with feature vectors encoding clinical variables and/or mutational features, and input a multilayer perceptron to predict patient risk. c: Deep learning Cox model to train a Cox proportional hazard model using patient-level image features extracted from the Inception V3 transfer learning architecture.
Key results
1. Hierarchical analysis of the risk of colon adenocarcinoma by pure image model and integrative model
The authors first classified patients as high risk (HR, OS < 3 years, N = 38), moderate risk (MR, 3 years< OS < 5 years, N = 45), and low risk (LR, OS > 5 years, N = 25) based on overall survival. Using HR and LR patients as a binary training set, a pure image model or an ensemble model was trained to predict the risk of colon adenocarcinoma risk, and the model was evaluated by KM survival curve and AUC value. The results showed that the image-only model was able to distinguish between HR and LR patients (AUC = 0.81±0.08), and that patients with HR and LR had significantly different survival curves (Figure 1a, b). However, when MR patients are added to the test set, the interval between survival curves decreases (Figure 1c).
Next, the authors compared the image-only model with a model based on clinical variables and/or mutational states and a comprehensive model combining WSIs, clinical variables, and mutational states (image & clinical & mutation models), and showed that the image-only model outperformed the clinical variable-only model (AUC = 0.71±0.12) or mutation-only model (AUC = 0.66±0.12), and the ensemble model combining clinical and mutational information (AUC = 0.69± 0.11), The fully integrated model behaves similarly to the image-only model in separating HR and LR patients (Figure 2a). KM survival curves show that an integrative model using only two data types (image & clinical model and image and mutation model, Figure 1d, e) is inferior to image & clinical & mutation model (Figure 1f).
Figure 1 Evaluation of the layering effect of multiple models
2. WSIs forecast heat maps reveal risk-related patterns
The authors used predictive heatmaps generated by image & clinical & mutation models to gain insight into the underlying morphology of CNNs associated with risk. The prediction heatmap shows the risk probability for each block predicted by the CNN, and pathologist review shows that nuclear shape, nuclear size pleomorphism, dense cellularity, and abnormal structure are indications of high risk (Figure 2). Accurate identification of tumor regions within WSI is a critical preliminary step affecting risk classification, and to test whether pathologist annotations for tumor regions can be replaced by computational methods, the authors used 228 pathologist annotations from independent WSI to construct computational tumor detectors. This detector shows high accuracy (Figure 3a, AUC >92%). KM curves of the image & clinical & mutation model using the computational tumor detector as input data show a clear separation between the high-risk and low-risk curves (Figure 3b).
Figure 2 Predictive heat map analysis
Figure 3 Establishment and verification of computational tumor detector
3. Model validation
Tumor regions (N=123) were collected and annotated in a separate COAD dataset from WSU and patients were stratified as HR (N=17), LR (N=97), or MR (N=9). In a validation set that included all HR, LR, and MR patients, the KM curve showed that the image-only model could not stratify high- and low-risk patients for this test set (Figure 4a), that only the clinical model provided statistically significant but modest stratification (Figure 4b), while the image+clinical model provided better patient cohort separation (Figure 4c). In the validation set for HR- and LR-only patients, it was found that all models were layered better than HR/MR/LR, and that the image and clinical models had better performance than the image-only and clinical-only models (Figure 4d–f). To test whether the model trained in TCGA-COAD is suitable for READ, the authors conducted model testing in the TCGA-READ validation queue. The results show that the image-only model successfully separates HR and LR patients (Figure 5a), but cannot stratify patients when MR patients are included in the test set (Figure 5b), while both the image+clinical model and the image-& clinical-mutant model can separate the patients (Figure 5c, d).
Figure 4 Model validation in the WSU-COAD dataset
Figure 5 Model testing in the TCGA-READ validation queue
4. Feature importance analysis in the model
In order to improve the interpretability of the deep learning model, the authors used SHAP to measure the contribution of each clinical or Inception v3 image feature to the model output in the TCGA training model and the WSU validation model, and found that T-phase, M-phase, and age were the most influential features in the ensemble model, although only two InceptionV3 features had comparable importance to these clinical variables, the total importance of InceptionV3 features (11.84) was higher than that of clinical variables ( 6.63) (Figure 6).
Figure 6 Feature importance analysis
brief summary
This paper uses public pathology image data, mutation data, and clinical variables to build a deep learning model to stratify tumors. Whether it is the analysis method of deep learning, or the selection of image and mutation data, it is to break the convention and seek innovation, so that the article can be sent to 11 points +! If you want to do innovative analysis and want to send a high-score student letter, this idea is quite good, don't miss it!
If you are still worried that there is no idea in the analysis of biologics, or if the analysis method is too simple and old-fashioned, you want innovative ideas, or if you are interested in single-cell analysis, multi-omics joint analysis and other directions, please contact Xiaoyun!
Xiaoyun continues to bring you the latest ideas of Shengxin, and more innovative analysis ideas please click the link. Friends who need to reproduce or learn more about analysis ideas are welcome to call Xiaoyun, and Xiaoyun is waiting for you in the wind and rain!
1. Q1 area 11 points + radiomics! Radiomics model construction + gene prognosis feature verification, break through the sharp weapon of life letter involution, and grasp to learn!
2. Receive in 1 month, unusual path of life letter + public pathological image analysis! Focusing on "immune hot and cold tumors", the ultra-simple analysis scored 6 points +, which is really fragrant!
3. 11 points + clinical database mining good article! The joint analysis of 3 major databases, plus multi-omics analysis of transcription protein metabolism, is rich beyond imagination, and it is quick to watch!
4. 7 points + double disease pure life letter analysis is alive again! Integrate "machine learning" marker screening, single gene analysis and drug molecule docking, and if you upgrade your ideas, you will lose a lot if you don't learn!
5. 9 points + wet and dry combination! TIL cell single-cell data analysis, combined machine learning model construction and IHC verification, strength to create tumor immunity letter good article!