CT/lung/src/losses/segmentation.py

from __future__ import annotations

import torch
import torch.nn.functional as F


def dice_loss(logits: torch.Tensor, target: torch.Tensor, eps: float = 1e-6) -> torch.Tensor:
    prob = torch.sigmoid(logits)
    dims = tuple(range(1, prob.ndim))
    numerator = 2.0 * torch.sum(prob * target, dim=dims)
    denominator = torch.sum(prob + target, dim=dims).clamp_min(eps)
    return 1.0 - torch.mean((numerator + eps) / (denominator + eps))


def segmentation_loss(logits: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
    return dice_loss(logits, target) + F.binary_cross_entropy_with_logits(logits, target.float())