Skip to main content

Table 2 Voxel-level NSCLC classification performance. Mean values and 95% confidence intervals are provided. Calibration metrics were not computable for GradCAM, integrated gradients and occlusion sensitivity methods. The clinician preference rate was calculated as the frequency with which clinicians preferred the method in 100 test instances, excluding instances in which they considered no model informative. DenseNet predictions were not included in the clinician preference test

From: Weakly supervised segmentation models as explainable radiological classifiers for lung tumour detection on CT images

Partition

Method

Precision

Recall

Dice

ECE

AUPR

Clinician preference rate

Validation

WSUnet

0.77 [0.75–0.8]

0.33 [0.29–0.36]

0.43 [0.39–0.46]

0.02 [0.01–0.02]

0.53 [0.49–0.56]

NA

Validation

sCNN GradCAM (16, 16, 64)

NA

NA

NA

NA

0.28 [0.25–0.3]

NA

Validation

sCNN GradCAM (8, 8, 128)

NA

NA

NA

NA

0.27 [0.25–0.29]

NA

Validation

sCNN integrated gradients

NA

NA

NA

NA

0.1 [0.09–0.1]

NA

Validation

sCNN occlusion sensitivity

NA

NA

NA

NA

0.04 [0.03–0.04]

NA

Validation

DenseNet GradCAM (16, 16, 64)

NA

NA

NA

NA

0.19 [0.17–0.21]

NA

Validation

DenseNet GradCAM (8, 8, 128)

NA

NA

NA

NA

0.23 [0.21–0.25]

NA

Validation

DenseNet occlusion sensitivity

NA

NA

NA

NA

0.03 [0.02–0.03]

NA

Test

WSUnet

0.78 [0.76–0.81]

0.24 [0.22–0.25]

0.33 [0.32–0.35]

0.01 [0.01–0.02]

0.4 [0.38–0.41]

0.72 [0.68–0.77]

Test

sCNN GradCAM (16, 16, 64)

NA

NA

NA

NA

0.36 [0.34–0.37]

0.2 [0.16–0.24]

Test

sCNN GradCAM (8, 8, 128)

NA

NA

NA

NA

0.23 [0.21–0.24]

0.05 [0.03–0.08]

Test

sCNN integrated gradients

NA

NA

NA

NA

0.11 [0.1–0.11]

0.01 [0.0–0.03]

Test

sCNN occlusion sensitivity

NA

NA

NA

NA

0.03 [0.03–0.03]

0.0 [0.0–0.01]

Test

DenseNet GradCAM (16, 16, 64)

NA

NA

NA

NA

0.13 [0.12–0.14]

NA

Test

DenseNet GradCAM (8, 8, 128)

NA

NA

NA

NA

0.23 [0.22–0.25]

NA

Test

DenseNet occlusion sensitivity

NA

NA

NA

NA

0.02 [0.02–0.02]

NA

  1. ECE Expected calibration error, AUPR Area under precision recall curve