jamesdolezal
commented
1 year ago Hi Itzg,
There are several things that can cause a model to fail to converge. This includes technical problems (e.g. issues with data or labels), suboptimal parameters (e.g. tile size, model architecture, hyperparameters), insufficient data (too few slides), or biological issues (e.g., the outcome can't be accurately predicted because there isn't sufficient true biological signal).
When troubleshooting models, the first place I like to start is by inspecting the dataset. Review the TFRecord tile extraction report and make sure that the data looks reasonable (tiles taken from inside areas of tumor). Then, use Dataset.summary() to quickly inspect the dataset to make sure that the TFRecords are valid and you have enough data for each outcome (0 and 1).
Can you paste the results of the following code, which will help me better understand the structure of your data?
import slideflow as sf
P = sf.load_project('.')
dataset = P.dataset(tile_px=256, tile_um=256)
dataset.summary()
ltzg
Dear Author,
I used the TCGA colorectal cancer WSI file to establish a 3-year recurrence prediction model (0 and 1) for colorectal cancer patients after surgery. I delineated the ROI of the WSI and homogenized the patches. I tried using different CNN networks (densenet, Restnet) to build the model, but the AUC was always 0.5. It seems that all the patches are classified as 0 or 1. Is there a better adjustment suggestion? Here is my code. ####################### import slideflow as sf
P = sf.create_project(
root='.',
annotations="./RFS1.csv",
slides='./slides1/',
)
P.extract_tiles(tile_px=256, tile_um=256)
P = sf.load_project('.')###每次退出都要重新加载项目文件
from slideflow.model import build_feature_extractor
P = sf.load_project('.')###每次退出都要重新加载项目文件
dataset = P.dataset(tile_px=256, tileum=256)
labels, = dataset.labels('category')
train_dts, val_dts = dataset.split(
model_type='categorical',
labels=labels,
val_strategy='k-fold',
val_k_fold=2,
k_fold_iter=1
)
hp = sf.ModelParams(
tile_px=256,
tile_um=256,
model='densenet',
)
Train with 5-fold cross-validation
P.train(
'category',
params=hp,
val_k_fold=2
)