physics_based_augmentation/flow_model/model/trainers/MR_signal_model.py

from typing import Tuple

import torch
import torch.nn as nn

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


#def denormalize_per_image(img_tensor, min_vals, max_vals):

    #min_t = torch.tensor(min_vals, dtype=img_tensor.dtype, device=img_tensor.device).view(3, 1, 1)
    #max_t = torch.tensor(max_vals, dtype=img_tensor.dtype, device=img_tensor.device).view(3, 1, 1)
    #img_denorm = img_tensor * (max_t - min_t) + min_t

    #return img_denorm


def denormalize_per_image(img_tensor, min_vals, max_vals):
    img_denorm = torch.zeros_like(img_tensor)
    img_tensor_c = img_tensor.clone()
    img_denorm[:, 0] = img_tensor_c[:, 0] * (max_vals[0] - min_vals[0]) + min_vals[0]
    img_denorm[:, 1] = img_tensor_c[:, 1] * (max_vals[1] - min_vals[1]) + min_vals[1]
    img_denorm[:, 2] = img_tensor_c[:, 2] * (max_vals[2] - min_vals[2]) + min_vals[2]

    return img_denorm


class MRSignalModel(nn.Module):

    def __init__(self) -> None:
        super().__init__()

    def forward(self, q_map, max_vals_img):

        t1_t2 = q_map[:, 0:2, :, :]
        t1 = t1_t2[:, 0:1, :, :]
        t2 = t1_t2[:, 1:2, :, :]
        pd = q_map[:, 2:3, :, :]

        TE_t1 = torch.tensor(0.00765).unsqueeze(-1).unsqueeze(-1).to(device)
        TR_t1 = torch.tensor(0.7).unsqueeze(-1).unsqueeze(-1).to(device)
        ESP_t1 = torch.tensor(0.00765).unsqueeze(-1).unsqueeze(-1).to(device)

        TE_t2 = torch.tensor(0.0765).unsqueeze(-1).unsqueeze(-1).to(device)
        TR_t2 = torch.tensor(6.0).unsqueeze(-1).unsqueeze(-1).to(device)
        ESP_t2 = torch.tensor(0.00765).unsqueeze(-1).unsqueeze(-1).to(device)

        TE_pd = torch.tensor(0.00669).unsqueeze(-1).unsqueeze(-1).to(device)
        TR_pd = torch.tensor(6.0).unsqueeze(-1).unsqueeze(-1).to(device)
        ESP_pd = torch.tensor(0.00669).unsqueeze(-1).unsqueeze(-1).to(device)

        eps = 1e-8
        denom_t2 = t2 + eps
        denom_t1 = t1 + eps

        etl = 10.0


        t1_w = torch.exp(-TE_t1 / denom_t2) * (1 - torch.exp(-(TR_t1 - etl * ESP_t1) / denom_t1))
        t2_w = torch.exp(-TE_t2 / denom_t2) * (1 - torch.exp(-(TR_t2 - etl * ESP_t2) / denom_t1))
        pd_w = torch.exp(-TE_pd / denom_t2) * (1 - torch.exp(-(TR_pd - etl * ESP_pd) / denom_t1))

        #return (
                #(torch.exp(log_pd)
                 #* t1_w),
                #(torch.exp(log_pd)
                 #* t2_w),
                #(torch.exp(log_pd)
                 #* pd_w)
            # 1600 #55000
        #)
        t1_weigh = (pd * t1_w) * 1600 #1500 #max_vals_img[:, 0] #1550 #55000

        t2_weigh = (pd * t2_w) * 1600 #1600 #max_vals_img[:, 1] #1660 #55000

        pd_weigh = (pd * pd_w) * 1600 #1600 #max_vals_img[:, 2] #2060 #55000

        return torch.cat((t1_weigh, t2_weigh, pd_weigh), 1)