代码优化：在医学图像分割模型评估中计算TN值

医学图像分割模型评估中的TN计算

以下代码在医学图像分割模型评估过程中计算TN值。该代码基于TP、FP和FN的计算基础，并使用 slicetotal 和已计算的TP、FP、FN值来计算TN。

**原始代码：**pythondef eval_net(net, dataset, slicetotal, batch_size=12, gpu=True): '''Evaluation without the densecrf with the dice coefficient'''

net.eval()    start = time.time()    dice_ = torch.zeros(14).cuda()    jac_ = torch.zeros(14).cuda()    NE = torch.zeros(14).cuda()    JNE = torch.zeros(14).cuda()

accuracy_ = torch.zeros(14).cuda()    precision_ = torch.zeros(14).cuda()    recall_ = torch.zeros(14).cuda()    specificity_ = torch.zeros(14).cuda()

print(1)    with torch.no_grad():        for i, b in enumerate(batch(dataset, batch_size)):

        imgs = np.array([k[0] for k in b]).astype(np.float32)            true_masks = np.array([k[1] for k in b])

        imgs = torch.from_numpy(imgs)            imgs = imgs.unsqueeze(1)            true_masks = torch.from_numpy(true_masks)

        pre_masks_eval = torch.zeros(true_masks.shape[0],14,256,256)            true_masks_eval = torch.zeros(true_masks.shape[0],14,256,256)            batchshape = true_masks.shape[0]

        batch_dice = torch.zeros(14).cuda()            if gpu:                imgs = imgs.cuda()                true_masks = true_masks.cuda()                net.cuda()

        output_img = net(imgs)            input = output_img.cuda()            pre_masks = input.max(1)[1].float() #索引代表像素所属类别的数字            for ak in range(14):                if ak == 0:                    continue                pre_masks_eval[:,ak] = (pre_masks==ak)                true_masks_eval[:,ak] = (true_masks==ak)                premasks = pre_masks_eval[:,ak].view(true_masks.shape[0],-1)                truemasks = true_masks_eval[:,ak].view(true_masks.shape[0],-1)

            intersection = premasks * truemasks                TP = intersection.sum(1)                FP = premasks.sum(1) - TP                FN = truemasks.sum(1) - TP

            for bk in range(true_masks.shape[0]):                    if TP[bk] == 0 and FP[bk] == 0 and FN[bk] == 0:                        NE[ak] += 1                        JNE[ak] += 1                    else:                        batch_dice[ak] = batch_dice[ak] + 2*TP[bk] / (2*TP[bk] + FP[bk] + FN[bk])                        jac_[ak] = jac_[ak] + TP[bk] / (TP[bk] + FP[bk] + FN[bk])

        dice_ = dice_ + batch_dice

    for knum in range(14):            dice_[knum] = dice_[knum] / (slicetotal - NE[knum])            jac_[knum] = jac_[knum] / (slicetotal - JNE[knum])    end = time.time()    print('time used:',end - start)

return dice_,jac_

**优化后的代码：**pythondef eval_net(net, dataset, slicetotal, batch_size=12, gpu=True): '''Evaluation without the densecrf with the dice coefficient'''

net.eval()    start = time.time()    dice_ = torch.zeros(14).cuda()    jac_ = torch.zeros(14).cuda()    NE = torch.zeros(14).cuda()    JNE = torch.zeros(14).cuda()

accuracy_ = torch.zeros(14).cuda()    precision_ = torch.zeros(14).cuda()    recall_ = torch.zeros(14).cuda()    specificity_ = torch.zeros(14).cuda()

TN = torch.zeros(14).cuda()  # 添加TN的初始化

print(1)    with torch.no_grad():        for i, b in enumerate(batch(dataset, batch_size)):

        imgs = np.array([k[0] for k in b]).astype(np.float32)            true_masks = np.array([k[1] for k in b])

        imgs = torch.from_numpy(imgs)            imgs = imgs.unsqueeze(1)            true_masks = torch.from_numpy(true_masks)

        pre_masks_eval = torch.zeros(true_masks.shape[0],14,256,256)            true_masks_eval = torch.zeros(true_masks.shape[0],14,256,256)            batchshape = true_masks.shape[0]

        batch_dice = torch.zeros(14).cuda()            if gpu:                imgs = imgs.cuda()                true_masks = true_masks.cuda()                net.cuda()

        output_img = net(imgs)            input = output_img.cuda()            pre_masks = input.max(1)[1].float() #索引代表像素所属类别的数字            for ak in range(14):                if ak == 0:                    continue                pre_masks_eval[:,ak] = (pre_masks==ak)                true_masks_eval[:,ak] = (true_masks==ak)                premasks = pre_masks_eval[:,ak].view(true_masks.shape[0],-1)                truemasks = true_masks_eval[:,ak].view(true_masks.shape[0],-1)

            intersection = premasks * truemasks                TP = intersection.sum(1)                FP = premasks.sum(1) - TP                FN = truemasks.sum(1) - TP

            TN[ak] = (slicetotal - NE[ak]) - (TP[ak] + FP[ak] + FN[ak])  # 计算TN值

            for bk in range(true_masks.shape[0]):                    if TP[bk] == 0 and FP[bk] == 0 and FN[bk] == 0:                        NE[ak] += 1                        JNE[ak] += 1                    else:                        batch_dice[ak] = batch_dice[ak] + 2*TP[bk] / (2*TP[bk] + FP[bk] + FN[bk])                        jac_[ak] = jac_[ak] + TP[bk] / (TP[bk] + FP[bk] + FN[bk])

        dice_ = dice_ + batch_dice

    for knum in range(14):            dice_[knum] = dice_[knum] / (slicetotal - NE[knum])            jac_[knum] = jac_[knum] / (slicetotal - JNE[knum])    end = time.time()    print('time used:',end - start)

return dice_,ja