Deep Learning Image Prediction

Advanced deep learning prediction module for histopathological image analysis. Implements Vision Transformer (ViT), ResNet, and other CNN architectures for biomarker prediction from H&E-stained tissue images.

Usage

deeplearningprediction(
  data,
  image_path_var,
  target_var,
  model_type = "vision_transformer",
  prediction_type = "classification",
  patch_size = 384,
  batch_size = 8,
  validation_split = 0.2,
  learning_rate = 1e-04,
  epochs = 10,
  attention_maps = TRUE,
  data_augmentation = TRUE,
  gray_zone_width = 5,
  thresholds = "",
  pretrained_weights = TRUE,
  freeze_backbone = FALSE,
  dropout_rate = 0.1,
  weight_decay = 1e-04,
  early_stopping = TRUE,
  save_model = FALSE,
  model_path = "",
  gpu_acceleration = TRUE,
  cross_validation = FALSE,
  cv_folds = 5,
  class_weights = TRUE,
  confidence_threshold = 0.8
)

Arguments

data

the data as a data frame

image_path_var

variable containing file paths to histological images

target_var

dependent variable to predict (biomarker status, classification)

model_type

deep learning architecture for image analysis

prediction_type

type of prediction task and output format

patch_size

size of image patches for processing (pixels)

batch_size

number of images processed simultaneously

validation_split

proportion of data used for validation

learning_rate

learning rate for model training

epochs

number of training epochs

attention_maps

create attention/saliency maps for model explainability

data_augmentation

apply image transformations during training

gray_zone_width

width of gray zone for cumulative logit models

thresholds

custom thresholds for ordinal prediction (comma-separated numbers)

pretrained_weights

initialize with ImageNet pretrained weights

freeze_backbone

freeze pretrained backbone for fine-tuning

dropout_rate

dropout rate for regularization

weight_decay

L2 regularization weight decay

early_stopping

stop training when validation loss stops improving

save_model

save trained model for future use

model_path

file path to save trained model

gpu_acceleration

utilize GPU for faster training if available

cross_validation

perform k-fold cross-validation

cv_folds

number of cross-validation folds

class_weights

apply class balancing weights for imbalanced datasets

confidence_threshold

minimum confidence threshold for predictions

Value

A results object containing:

results$modelSummary					a html
results$dataInfo					Summary of input data and preprocessing
results$trainingMetrics$trainingHistory					Epoch-by-epoch training progress
results$trainingMetrics$finalMetrics					Overall model performance assessment
results$predictionResults$classificationResults					a table
results$predictionResults$cumulativeLogitResults					a table
results$performanceAnalysis$confusionMatrix					Classification confusion matrix
results$performanceAnalysis$classificationReport					Precision, recall, F1-score per class
results$performanceAnalysis$rocAnalysis					ROC AUC scores for each class
results$explainabilityAnalysis$attentionSummary					Summary of attention patterns across samples
results$explainabilityAnalysis$featureImportance					Most important image features for prediction
results$crossValidationResults$cvMetrics					Performance across CV folds
results$crossValidationResults$cvSummary					Mean and standard deviation of CV metrics
results$trainingCurvePlot					Training and validation loss/accuracy over epochs
results$confusionMatrixPlot					Heatmap of confusion matrix
results$rocCurvePlot					ROC curves for each class
results$attentionHeatmapPlot					Sample attention maps showing model focus areas
results$predictionDistributionPlot					Distribution of prediction confidence scores
results$cumulativeLogitPlot					Cumulative probability plots for ordinal prediction
results$crossValidationPlot					Box plots of CV performance metrics
results$modelArchitecturePlot					Visual representation of the neural network architecture

Tables can be converted to data frames with asDF or as.data.frame. For example:

results$dataInfo$asDF

as.data.frame(results$dataInfo)

Examples

# Basic Vision Transformer prediction
deeplearningprediction(
    data = image_data,
    image_path_var = "image_file_path",
    target_var = "biomarker_status",
    model_type = "vision_transformer",
    attention_maps = TRUE,
    validation_split = 0.2
)

# Cumulative logit with gray-zone
deeplearningprediction(
    data = ordinal_data,
    image_path_var = "image_paths",
    target_var = "ki67_score",
    model_type = "vision_transformer",
    prediction_type = "cumulative_logit",
    gray_zone_width = 5.0,
    thresholds = c(5, 10, 15, 20)
)