Table 3 Comparative analysis of Tiger Model against other data augmentation methods in the classification of benign and malignant rare subtypes of thyroid cancer

From: Improving AI models for rare thyroid cancer subtype by text guided diffusion models

Augmentation Method

(Augmentation data

number = 30 k)

AUROC (95% CI)

Benign vs PTC (Resnet50)

Benign vs FTC (Resnet50)

Benign vs MTC (Resnet50)

Valid

Test

Valid

Test

Valid

Test

① No Data Augmentation

0.8532

(0.83–0.93)

0.8677

(0.83–0.92)

0.7312

(0.68–0.75)

0.7364

(0.69–0.75)

0.7438

(0.63–0.78)

0.7523

(0.67–0.79)

Basic Data Augmentation

② Basic image manipulation38

0.8763

(0.85–0.88)

0.8614

(0.84–0.87)

0.6999

(0.68–0.72)

0.6872

(0.65–0.69)

0.6988

(0.63–0.70)

0.6511

(0.63–0.66)

③ Mixing image39

0.8321

(0.81–0.84)

0.8563

(0.82–0.87)

0.6761

(0.65–0.69)

0.6321

(0.60–0.64)

0.6947

(0.64–0.72)

0.6827

(0.63–0.72)

④ Random erasing40

0.8426

(0.83–0.86)

0.8637

(0.82–0.89)

0.6590

(0.56–0.72)

0.6472

(0.62–0.73)

0.6743

(0.63–0.69)

0.7182

(0.65–0.72)

⑤ Feature space augmentation41

0.8532

(0.83–86)

0.8452

(0.81–0.87)

0.6842

(0.65–0.73)

0.6828

(0.60–0.72)

0.7083

(0.67–0.73)

0.7427

(0.72–0.75)

Deep Learning Generation

⑥ PG-GAN42

0.8427

(0.82–0.86)

0. 8453

(0.82–0.86)

0.7022

(0.68–0.75)

0.6732

(0.63–0.72)

0.7529

(0.68–0.81)

0.7468

(0.72–0.79)

⑦ Diffusion Transformers43

0.8743

(0.84–0.92)

0.8703

(0.85–0.92)

0.7380

(0.71–0.75)

0.7328

(0.69–0.75)

0.7121

(0.65–0.73)

0.7252

(0.69–0.76)

⑧ Imagen44

0.8538

(0.82–0.86)

0.8412

(0.84–0.86)

0.7543

(0.72–0.80)

0.7577

(0.68–0.82)

0.7574

(0.74–0.82)

0.7581

(0.63–0.79)

⑨ Stable Diffusion24

0.8694

(0.84–0.89)

0.8503

(0.85–0.89)

0.7471

(0.73–0.79)

0.7737

(0.76–0.82)

0.7627

(0.74–0.82)

0.7582

(0.73–0.76)

⑩ Stable Diffusion + ControlNet63

0.8473

(0.84–0.89)

0.8527

(0.83–0.87)

0.7033

(0.65−0.75)

0.7172

(0.69–0.74)

0.7261

(0.69–0.76)

0.7425

(0.71–0.76)

⑪ Tiger-N

0.8882

(0.84–0.89)

0.8969

(0.84–0.92)

0.8067

(0.79–0.82)

0.8132

(0.78–0.83)

0.7733

(0.76–0.83)

0.8027

(0.79–0.82)

⑫ Tiger-F

0.9127

(0.90–0.93)

0.9263

(0.85–0.94)

0.8338

(0.76–0.83)

0.8442

(0.75–0.85)

0.8043

(0.75–0.83)

0.8234

(0.78–0.85)

  1. ① refers to the baseline without any data augmentation. ② refers to Flipping, Cropping, Rotation, Translation, and Noise injection operations. ⑥⑦ refers to the type of image-to-image generation models. ⑧⑨⑩⑪⑫ refers to the type of Text-to-image generation models. ⑧⑨⑩⑪⑫ are fine-tuned using our thyroid image-level features. ⑪ is the Tiger Model trained and generated solely using disease subtype names as prompts. ⑫ is a Tiger Model trained and generated using the nodule image’s detailed description text as prompts.
Back to article page