Artificial Intelligence-based Tumor Segmentation in Mouse Models of Lung Adenocarcinoma

Abstract

Background: Mouse models are highly effective for studying the pathophysiology of lung adenocarcinoma and evaluating new treatment strategies. Treatment efficacy is primarily determined by the total tumor burden measured on excised tumor specimens. The measurement process is time-consuming and prone to human errors. To address this issue, we developed a novel deep learning model to segment lung tumor foci on digitally scanned hematoxylin and eosin (H&E) histology slides.

Methods: Digital slides of 239 mice from 9 experimental cohorts were split into training (n=137), validation (n=37), and testing cohorts (n=65). Image patches of 500×500 pixels were extracted from 5× and 10× magnifications, along with binary masks of expert annotations representing ground-truth tumor regions. Deep learning models utilizing DeepLabV3+ and UNet architectures were trained for binary segmentation of tumor foci under varying stain normalization conditions. The performance of algorithm segmentation was assessed by Dice Coefficient, and detection was evaluated by sensitivity and positive-predictive value (PPV).

Results: The best model on patch-based validation was DeepLabV3+ using a Resnet-50 backbone, which achieved Dice 0.890 and 0.873 on validation and testing cohort, respectively. This result corresponded to 91.3 Sensitivity and 51.0 PPV in the validation cohort and 93.7 Sensitivity and 51.4 PPV in the testing cohort. False positives could be reduced 10-fold with thresholding artificial intelligence (AI) predicted output by area, without negative impact on Dice Coefficient. Evaluation at various stain normalization strategies did not demonstrate improvement from the baseline model.

Conclusions: A robust AI-based algorithm for detecting and segmenting lung tumor foci in the pre-clinical mouse models was developed. The output of this algorithm is compatible with open-source software that researchers commonly use.

Figure 1

Example WSI and associated image tiles from training set. (Left) WSI with regions of tumor outlined in red and regions of any lung tissue, normal or malignant, outlined in blue. (Right-top) A representative 5× tile extracted from WSI and binary mask converted from expert annotations. The same patch transformed by Macenko and Vahadane are shown. (Right-bottom) A representative 10× tile extracted from WSI, representing the bottom right quadrant of 5× tile, and associated binary masks and normalized features.

Figure 2

Good Performance Cases for 5× DeepLabV3+ Model. (Left) WSI from Aperio scanner with Dice Coefficient 0.930 from validation set. (Right) WSI from Hamamatsu scanner with Dice Coefficient 0.960 from the test set. For ground-truth annotations, tumor regions are outlined in red and total lung regions are outlined in green. AI outputs are outlined in yellow.

Figure 3

Worst Performance Cases for 5× DeepLabV3+ Model. (Left) WSI from Aperio scanner with Dice Coefficient 0.778 from test set. (Right) WSI from Hamamatsu scanner with Dice Coefficient 0.227 from the test set. For ground truth annotations, tumor regions are outlined in red and total lung regions are outlined in green. AI outputs are outlined in yellow.

Figure 4

Negative Test Case for 5× DeepLabV3+ Model. (Left) Ground truth annotation demonstrating only lung regions, without the presence of tumor foci. (Right) AI produced a single false positive of approximately 100 μM × 20 μM in size, outlined in yellow.

Figure 5

FROC Curve for Training and Validation sets for 5× DeepLabV3+ Model. Risk is assessed by AI-predicted foci size demonstrating reduction of false positives per image by increasing cut-off threshold (shown in increments of 400 μM²).

Figure 6

5× DeepLabV3+ Model Bland‑Altman Plot for total tumor burden assessment by Expert vs AI for Validation and Testing datasets.