Performance Analysis of Artifact Detection in Digital Pathology using GrandQC
Implementation Plan:
————————–
Scenario 1: ( with transfer learning ):
———————————————-
Step 1: Initially, we collect and load data from H&E and IHC dataset(Using below mentioned 4 .svs files)
Step 2: Next, we preprocess the data into patches, normalize stain variations, and align with GeoJSON masks for artifact labeling.
Step 3: Next, we implement user defined choices such as GrandQC, HistoQC, PathProfiler and HistoROI to detect artifacts and generate output masks for selected images.
Step 4: Next, we train the selected QC’s output by comparing DICE scores with ground truth .
Step 5: Next, we apply transfer learning to adapt selected QC and then fine-tune the Hibou foundation model for artifact segmentation.
Step 6: Next, we analyze pixel-level, patch-level, and slide-level results.
Step 7: Finally, we plot performance for the following metrics:
7.1: Number of epochs vs. Accuracy (%)
7.2: Number of epochs vs. Precision (%)
7.3: Number of epochs vs. Recall (%)
7.4: Number of epochs vs. F1-score (%)
Scenario 2: ( without transfer learning ):
————————————————–
Step 1: Initially, we collect and load data from H&E and IHC dataset(Using below mentioned 4 .svs files)
Step 2: Next, we preprocess the data into patches, normalize stain variations, and align with GeoJSON masks for artifact labeling.
Step 3: Next, we implement user defined choices such as GrandQC, HistoQC, PathProfiler and HistoROI to detect artifacts and generate output masks for selected images.
Step 4: Next, we train the selected QC’s output by comparing DICE scores with ground truth .
Step 5: Next, we analyze pixel-level, patch-level, and slide-level results.
Step 6: Finally, we plot performance for the following metrics:
6.1: Number of epochs vs. Accuracy (%)
6.2: Number of epochs vs. Precision (%)
6.3: Number of epochs vs. Recall (%)
6.4: Number of epochs vs. F1-score (%)
Scenario 3: ( with Foundation Model):
———————————————-
Step 1: Initially, we collect and load data from H&E and IHC dataset(Using below mentioned 4 .svs files)
Step 2: Next, we preprocess the data into patches, normalize stain variations, and align with GeoJSON masks for artifact labeling.
Step 3: Next, we implement user defined choices such as GrandQC, HistoQC, PathProfiler and HistoROI to detect artifacts and generate output masks for selected images.
Step 4: Next, we train the selected QC’s output by comparing DICE scores with ground truth .
Step 5: Next, we perform artifact segmentation using a fine-tune Hibou foundation model.
Step 6: Next, we analyze pixel-level, patch-level, and slide-level results.
Step 7: Finally, we plot performance for the following metrics:
7.1: Number of epochs vs. Accuracy (%)
7.2: Number of epochs vs. Precision (%)
7.3: Number of epochs vs. Recall (%)
7.4: Number of epochs vs. F1-score (%)
Software requirement:
—————————
1. Development Tool: Python 3.11.4 or above
2. Operating System: Windows 10 (64-bit) or above
Dataset:
———-
Link :- http://histoqcrepo.com/
– RSNA_38_LA1_V.svs
– TCGA-C8-A12P-01Z-00-DX1.670B5DE8-07B0-4E4C-93FA-FA3DFFCCE50D.svs
– TCGA-D8-A1J9-01Z-00-DX2.E1C59487-9563-4501-845F-2067A0C5C59B.svs
– TCGA-D8-A141-01Z-00-DX2.DBD0D81E-28FC-4466-BDE3-94753BD6CBEB.svs
– TCGA-E9-A1R4-01Z-00-DX1.B04D5A22-8CE5-49FD-8510-14444F46894D.svs
Note:
——
1) If the plan does not meet your requirements, provide detailed steps, parameters, models, or expected results in advance. Once implemented, changes won’t be possible without prior input; otherwise, we’ll proceed as per our implementation plan.
2) If the plan satisfies your requirement, Please confirm with us.
3) Project based on Simulation only, not a real time project.
4) If you have any changes in the Dataset , kindly provide before implementation. Our work is completely based on dataset values.
5) Please understand that any modifications made to the confirmed implementation plan will not be made after the project development.
