Table 1.
Set of hand-crafted features used for comparison against the CNN based feature learning approach.
Fig 1.
Overview of the high-throughput adaptive sampling for whole-slide histopathology images method (HASHI) based on CNNs for automated detection of invasive breast cancer (BCa) in WSIs. Training data cohorts: Hospital of the Univ. of Pennsylvania (HUP) and Case Western Reserve Univ. (CWRU). Validation/Testing data cohorts: Cancer Institute of New Jersey (CINJ) and The Cancer Genome Atlas (TCGA).
Fig 2.
Illustration of the CNN architecture used to distinguish between invasive and non-invasive breast cancer (BCa) image tiles.
The architecture is a 2-layer CNN with 256 neurons in the first layer convolutional (C1) and subsampling/pooling layer (S2) and 256 neurons in the fully-connected layer (FC), (i.e. CS256-FC256). Amongst the various architectures considered, this architecture was selected because it has a good trade-off between classification performance and a shallower architecture (fewer layers).
Table 2.
Breast cancer data cohorts used for training, validation and testing in the experimental evaluation.
Table 3.
Comparison between CNN models and state-of-the-art hand-crafted features trained with D3 and evaluated on D4 in terms of AUC.
Fig 3.
Comparison between sampling methods (regular and dense) with HASHI using gradient-based quasi-Monte Carlo sampling (grad-qmc-halton) [59, 61].
The new unseen WSI (A) with its corresponding ground truth annotation from an expert pathologist (B). The probability maps using regular sampling with a step size equal to the patch size (C) and regular dense sampling with step size equal to 1 pixel (D). HASHI involves an iterative process of extracting patch samples (E-H) and obtaining the corresponding probability maps (I-L) for the 1st (E, I), 2nd (F, J), 8th (G, K) and 20th iteration respectively (H, L).
Fig 4.
Quantitative evaluation of the different sampling strategies in terms of average Dice coefficient (y-axis) versus the number of samples (x-axis) used.
All strategies were trained with D3 and evaluated with D5.
Fig 5.
GPU memory size requirements (Megabytes) for different image dimensions (height × width × channels) for the experimentation of FCN based on CNN2 model.
Table 4.
Invasive BCa detection performance of HASHI and the equivalent FCN architecture on D5 in terms of Dice coefficient.
Table 5.
Invasive BCa detection performance of our method on the Dtest testing dataset in terms of Dice, PPV, NPV, TPR, TNR, FPR, FNR.
Fig 6.
Performance comparison between HASHI and M1 in terms of Dice coefficient in the independent Dtest test data cohort by varying the classification threshold of the invasive BCa probability map.
Fig 7.
Results of the invasive BCa probability maps (second and fourth rows) predicted by HASHI on representative WSIs from Dtest compared to the ground truth annotations from expert pathologists (first and third rows).
Red regions represent locations identified by HASHI as having a high likelihood of cancer presence while the blue regions represent the lowest likelihood of cancer presence.