Skip to main content

Table 2 Classification performance of randomly initialized versus pre-trained embeddings

From: Multi-label annotation of text reports from computed tomography of the chest, abdomen, and pelvis using deep learning

Organ

Label

# Pos

Random Initialization (AUC)

Pre-trained (AUC)

Lungs/Pleura

Atelectasis

9329

0.9968 (0.9961–0.9974)

0.9973 (0.9967–0.9997)

Nodule

10,183

0.9913 (0.9904–0.9922)

0.9935 (0.9928–0.9943)

Emphysema

3659

0.9972 (0.9963–0.9982)

0.9980 (0.9972–0.9987)

Effusion

5625

0.9975 (0.9970–0.9980)

0.9984 (0.9980–0.9989)

Normal

3110

0.9990 (0.9985–0.9995)

0.9990 (0.9982–0.9997)

Liver/Gallbladder

Stone

1981

0.7849 (0.7739–0.7059)

0.9761 (0.9721–0.9801)

Lesion

6463

0.9675 (0.9646–0.9700)

0.9946 (0.9936–0.9955)

Dilatation

1497

0.8120 (0.8013–0.8228)

0.9926 (0.9906–0.9945)

Fatty

1795

0.9984 (0.9851–0.9917)

0.9991 (0.9986–0.9996)

Normal

3162

0.9745 (0.9716–0.9773)

0.9762 (0.9950–0.9974)

Kidneys/Ureters

Stone

2548

0.9562 (0.9514–0.9609)

0.9792 (0.9764–0.9819)

Atrophy

750

0.9523 (0.9436–0.9611)

0.9955 (0.9936–0.9973)

Lesion

4817

0.9757 (0.9731–0.9783)

0.9900 (0.9886–0.9915)

Cyst

4164

0.9862 (0.9843–0.9881)

0.9926 (0.9914–0.9939)

Normal

2048

0.9909 (0.9890–0.9928)

0.9980 (0.9980–0.9992)

  1. Classification performance of randomly initialized versus pre-trained embeddings for each disease. “# Pos” represents the number of positives for that label. Values are reported as area under the curve (AUC) with 95% confidence interval (CI). Bolded values represent an equivalent AUC or increase in performance