src/losses.py

from __future__ import annotations

import torch
import torch.nn as nn
import torch.nn.functional as F


def dice_loss_from_logits(logits: torch.Tensor, target: torch.Tensor, eps: float = 1.0) -> torch.Tensor:
    pred = torch.sigmoid(logits)
    dims = (1, 2, 3)
    inter = (pred * target).sum(dim=dims)
    denom = pred.sum(dim=dims) + target.sum(dim=dims)
    return (1.0 - (2.0 * inter + eps) / (denom + eps)).mean()


def soft_erode(x: torch.Tensor) -> torch.Tensor:
    vertical = -F.max_pool2d(-x, kernel_size=(3, 1), stride=1, padding=(1, 0))
    horizontal = -F.max_pool2d(-x, kernel_size=(1, 3), stride=1, padding=(0, 1))
    return torch.minimum(vertical, horizontal)


def soft_dilate(x: torch.Tensor) -> torch.Tensor:
    return F.max_pool2d(x, kernel_size=3, stride=1, padding=1)


def soft_open(x: torch.Tensor) -> torch.Tensor:
    return soft_dilate(soft_erode(x))


def soft_skeletonize(x: torch.Tensor, iterations: int = 20) -> torch.Tensor:
    skeleton = F.relu(x - soft_open(x))
    eroded = x
    for _ in range(max(int(iterations), 1)):
        eroded = soft_erode(eroded)
        opened = soft_open(eroded)
        delta = F.relu(eroded - opened)
        skeleton = skeleton + F.relu(delta - skeleton * delta)
    return skeleton.clamp(0, 1)


def cldice_loss_from_logits(logits: torch.Tensor, target: torch.Tensor, iterations: int = 20, eps: float = 1.0) -> torch.Tensor:
    pred = torch.sigmoid(logits).clamp(0, 1)
    target = target.clamp(0, 1)
    pred_skel = soft_skeletonize(pred, iterations)
    target_skel = soft_skeletonize(target, iterations)
    dims = (1, 2, 3)
    topology_precision = ((pred_skel * target).sum(dim=dims) + eps) / (pred_skel.sum(dim=dims) + eps)
    topology_sensitivity = ((target_skel * pred).sum(dim=dims) + eps) / (target_skel.sum(dim=dims) + eps)
    cldice = (2.0 * topology_precision * topology_sensitivity) / (topology_precision + topology_sensitivity + 1e-6)
    return (1.0 - cldice).mean()


class SegmentationLoss(nn.Module):
    def __init__(
        self,
        bce_weight: float = 1.0,
        dice_weight: float = 1.0,
        cldice_weight: float = 0.5,
        skeleton_iterations: int = 20,
    ):
        super().__init__()
        self.bce_weight = bce_weight
        self.dice_weight = dice_weight
        self.cldice_weight = cldice_weight
        self.skeleton_iterations = skeleton_iterations

    def forward(self, logits: torch.Tensor, target: torch.Tensor) -> tuple[torch.Tensor, dict[str, torch.Tensor]]:
        bce = F.binary_cross_entropy_with_logits(logits, target)
        dice = dice_loss_from_logits(logits, target)
        cldice = cldice_loss_from_logits(logits, target, self.skeleton_iterations)
        loss = self.bce_weight * bce + self.dice_weight * dice + self.cldice_weight * cldice
        return loss, {"loss": loss.detach(), "bce_loss": bce.detach(), "dice_loss": dice.detach(), "cldice_loss": cldice.detach()}